DE112020007659T5 - Robot collision detection device - Google Patents

Abstract

A robot collision detection device (1) includes driving torque calculation means (3) that calculates an estimated value of the driving torque of a robot, torque estimation error model training means (5) that learns a difference and a variation of the difference, the difference between the based on a motor current for driving the robot calculated drive torque and the estimated value calculated by the drive torque calculation means (3), threshold value calculation means (8) that calculate a threshold value based on a torque estimation error model trained by the torque estimation error model training means (5), and collision determination means (9) that make a determination about the occurrence of a Collision of the robot with an object by comparing the difference with the threshold value calculated by the threshold calculation means (8), the difference between the driving torque calculated based on the motor current for driving the robot and the estimated value calculated by the driving torque calculation means (3). .

Description

Area

The present disclosure relates to a robot collision detection device for detecting a collision of a robot with an object near the robot when the robot collides with the object.

General state of the art

A robot conventionally detects the contact or collision of the robot with a device or a worker in the vicinity of the robot, and stops its operation when the robot comes into contact with or collides with the device or worker to avoid damage to the device or a prevent injury to the worker near the robot and the robot itself. Conventionally, a robot controller calculates the torque value required for the operation to be performed by the robot and derives a measured torque value based on an actual motor current value or a measured value from a torque sensor provided in a drive mechanism unit. The robot controller compares the torque reading to the required torque value and determines that the robot has collided with a device or worker in the vicinity of the robot when the difference between the required torque value and the torque reading exceeds a threshold. In order to detect a collision of a robot with a device or a worker near the robot with high sensitivity, a technology is required to lower the threshold as much as possible without causing false detection.

Patent Literature 1 discloses a technology for estimating parameters related to an inertial force and frictional force of a robot and improving, using the estimated parameters, accuracy of calculation or accuracy of estimation of the torque required for the robot performed operation is required. Patent Literature 1 also discloses a technology for reducing the effect of a factor not modeled by a required torque calculation unit by using a high-pass filter.

citation list

patent literature

Patent Literature 1: Japanese Translation of PCT International Publication No. 2016-511699

abstract

Technical problem

The conventional technology can deal with a variation of a parameter related to the dynamics of a robot, such as an effect of a change in a workpiece held by a hand attached to an arm end of the robot, but takes an effect into account the variation in torque during robot operation under the same conditions. That is, during an operation with a large torque variation, the conventional technology may cause misdetection of the occurrence of a collision even when the robot has not collided with a device or a worker near the robot. In addition, the conventional technology uses a high-pass filter to cancel the effect of an unmodeled factor such as an effect of elasticity or non-linearity of friction caused by the gear mechanism. This can still allow a collision to be overlooked, which is difficult to detect just considering errors in a high-frequency region, when the collision occurs in such a way that, for example, the robot slowly approaches a device or a worker near the robot meets.

The present disclosure was developed in view of the foregoing, and an object of the present disclosure is to provide a robot collision detection device that detects a collision of a robot with an object near the robot with relatively high sensitivity when the robot collides with the object. while accounting for the effect of an unmodeled factor while preventing misdetection.

the solution of the problem

In order to solve the problem and achieve the object described above, a robot collision detection device according to the present disclosure includes driving torque calculation means that calculates an estimated value of driving torque of a robot, torque estimation error model training means that calculates a difference and a variation of the difference learn, wherein the difference between the drive torque calculated on the basis of a motor current for driving the robot or the drive torque derived from a torque sensor provided on a drive unit and the estimated value calculated by the drive torque calculation means, threshold value calculation means that sets a threshold value on the based on a torque estimation error model trained by the torque estimation error model training means, and collision determination means that makes a determination on the occurrence of a collision of the robot with an object by comparing the difference with the threshold value calculated by the threshold calculation means, the difference between the based on the motor current for driving of the robot calculated drive torque and the estimated value calculated by the drive torque calculation means.

Advantageous Effects of the Invention

A robot collision detection device according to the present disclosure provides an advantage of being able to detect a collision of a robot with an object in the vicinity of the robot with relatively high sensitivity when the robot collides with the object while the effect of a non-modeled factor, including a variation thereof, is taken into account while preventing false detection.

character list

• 1 12 is a diagram illustrating a configuration of a robot collision detection device according to a first embodiment.
• 2 12 is a diagram illustrating a configuration of a robot collision detection device according to a second embodiment.
• 3 12 is a diagram illustrating a configuration of a robot collision detection device according to a fifth embodiment.
• 4 12 is a diagram illustrating a configuration of a robot collision detection device according to a sixth embodiment.
• 5 12 is a diagram illustrating a configuration of a robot collision detection device according to an eighth embodiment.
• 6 12 is a diagram illustrating a configuration of a robot collision detection device according to a ninth embodiment.
• 7 12 is a diagram illustrating a processor when at least a part of the driving torque calculation means, operating state calculation means, torque estimation error model training means, conversion means, error calculation means, threshold value calculation means, and collision determination means included in the robot collision detection device according to the first embodiment is implemented in the processor.
• 8th 12 is a diagram illustrating a processing circuit when at least a part of the driving torque calculation means, the operating state calculation means, the torque estimation error model training means, the conversion means, the error calculation means, the threshold value calculation means, and the collision determination means included in the robot collision detection device according to the first embodiment is implemented in the processing circuit .

Description of the embodiments

A robot collision detection device according to embodiments will be described in detail below with reference to the drawings.

First embodiment.

1 12 is a diagram illustrating a configuration of a robot collision detection device 1 according to a first embodiment. The following description may refer to the robot collision detection device 1 as the “collision detection device 1”. The collision detection device 1 includes a robot control device 2 that controls a robot. The robot is in 1 not illustrated.

The robot controller 2 includes driving torque calculation means 3 which receives information representing the motor position along each axis for driving the robot and calculates an estimated value of the driving torque of the robot. The driving torque calculation means 3 differentiates the motor position to calculate the motor speed and the motor acceleration along each axis. The Driving torque calculation means 3 calculates the driving torque of the robot based on the motor position, motor speed and motor acceleration and the equation of motion (1) of the robot as described below. $$\mathrm{\tau }=\text{M}\left(\text{q}\right)\text{a}+\text{H}\left(\text{q}\text{, v}\right)+\text{G}\left(\text{q}\right)+\text{f}\left(\text{v}\right)$$

In Equation (1), "τ" is a vector containing the driving torque along each axis of the robot, and "q" is a vector containing the position along each axis of the robot, which is equivalent to the motor position along each axis of the robot is shown in the form of an output position of the gear mechanism. The parameter "a" is a vector containing the acceleration along each axis of the robot, which is equivalent to the motor acceleration and is represented in the form of an output acceleration of the gear mechanism along each axis of the robot, and "v" is a vector containing the speed along each axis of the robot, which is equivalent to the motor speed and is represented in terms of an output speed of the gear mechanism along each axis of the robot.

The term M(q)a represents the inertial force along each axis of the robot. The term h(q,v) represents the Coriolis centrifugal force. The term g(q) represents gravity. The term f(v) represents represents the frictional force. These terms are components of the drive torque. The frictional force is a sum of Coulomb frictional force versus velocity direction and viscous frictional force versus velocity direction and magnitude. The simplest viscous friction force model is a velocity proportional model, and the first embodiment uses a velocity proportional viscous friction model. The viscous frictional force model may be a velocity polynomial model or a velocity exponential model.

The collision detection device 1 further includes operational state calculation means 4 arranged outside the robot control device 2. The operational state calculation means 4 calculates a state quantity related to the operational state of the robot. More specifically, the operation state calculation means 4 calculates a state quantity including all or a part of the motor speed information, the motor acceleration information and a part of the drive torque components of the robot. The expression “part of the components of the driving torque of the robot” refers to one of the inertial force M(q)a, the Coriolis centrifugal force h(q,v), the gravitational force g(q) and the frictional force f(v) or one Sum of drive torque components, such as M(q)a+h(q,v). In the first embodiment, the operating state calculation means 4 receives information representing the motor position, differentiates the motor position to calculate the motor speed, and calculates an operating state quantity representing the speed along each axis of the robot, which is equivalent to the motor speed, and forms an output speed of the gear mechanism along each Axis of the robot is shown.

The collision detection device 1 further includes torque estimation error model training means 5 arranged outside the robot control device 2. The torque estimation error model training means 5 learns a difference and a variation of the difference, the difference between the driving torque calculated based on a motor current for driving the robot and an estimated value calculated by the driving torque calculation means 3 of the drive torque is present. The operating state calculation means 4 outputs the information representing the operating state quantity to the torque estimation error model training means 5 . The torque estimation error model training means 5 performs learning using the state quantity calculated by the operating state calculation means 4 as an input signal for a correction function. One or both of the operating state calculation means 4 and the torque estimation error model training means 5 may be arranged within the robot control device 2 .

The robot control device 2 further includes conversion means 6 which calculate the measured torque. The measured torque is a measured value of the driving torque along each axis of the robot, which is obtained by converting the motor current based on the torque constant of the motor and the speed ratio of the gear mechanism. The collision detection device 1 further includes error calculation means 7 arranged outside the robot control device 2. The driving torque calculation means 3 outputs information representing the estimated torque to the error calculation means 7. The estimated torque is an estimated value of the driving torque. The conversion means 6 outputs information representing the measured torque to the error calculation means 7 . The Error calculating means 7 subtracts the estimated torque calculated by the driving torque calculating means 3 from the measured torque obtained from the converting means 6 to calculate a torque estimation error which is a difference between the measured torque and the estimated torque. The error calculation means 7 outputs the information representing the torque estimation error to the torque estimation error model training means 5 .

When performing the learning, the torque estimation error model training means 5 receives the information representing the operation state quantity output from the operation state calculation means 4 and the information representing the torque estimation error and output from the error calculation means 7 . The torque estimation error model training means 5 includes training means 5a using a non-parametric technique for each axis. The training means 5a use Gaussian process regression. The training means 5a, which are of a non-parametric type, can also use kernel density estimation or a k-nearest neighbor algorithm.

Regarding the learning performed by the torque estimation error model training means 5, let x1, x2, ..., xn be the output from the operating state calculation means 4, respectively, and yi be the difference between the measured torque corresponding to xi and the estimated torque calculated by the Driving torque calculating means 3, where n is an integer greater than or equal to 2 and i is an integer ranging from 1 to n. The torque estimation error model training means 5 performs learning over a hyperparameter of a kernel function used in Gaussian process regression using n pairs each with an input of x and an output of y, i.e. H. D={(x1,y1), (x2,y2), ..., (xn,yn)} , as learning data. Popular kernel functions for Gaussian process regression include Gaussian and radial basis function (RBF) kernels. The torque estimation error model training means 5 uses an RBF as a kernel function. The torque estimation error model training means 5 can also use a kernel function other than an RBF, e.g. B. an exponential kernel or a periodic kernel.

After the completion of the learning, the torque estimation error model training means 5 obtains a torque estimation error model for calculating a predicted distribution of y* based on a new input x*. Equation (2) below gives an example of the torque estimation error model. $$\text{P}\left(\text{y*}|\text{x*}\text{,D}\right)=\text{N}\left({\text{k*}}^{\text{T}}{\text{K}}^{-1}\text{y}\text{,k**}-{\text{k*}}^{\text{T}}{\text{K}}^{-1}\text{k*}\right)$$

$$\text{k**}=\text{k}\left(\text{x*}\text{,x*}\right)$$

k() ist eine Kernel-Funktion. K ist eine Matrix, die als Kernel-Matrix bezeichnet wird, deren ij-tes Element k(xi, xj) ist, wobei j eine ganze Zahl im Bereich von 1 bis n ist. N(b, σ²) ist eine Wahrscheinlichkeitsdichtefunktion einer Gauß-Verteilung mit einem Mittelwert von b und einer Varianz von σ².The k* factor of Equation (2) is expressed by Equation (3) below, and the k** factor of Equation (2) is expressed by Equation (4) below. $$\text{k*}={\left(\text{k}\left(\text{x*}\text{,x1}\right),\text{k}\left(\text{x*}\text{,x2}\right),...,\text{k}\left(\text{x*}\text{,xn}\right)\right)}^{\text{T}}$$

$$\text{k**}=\text{k}\left(\text{x*}\text{,x*}\right)$$

k() is a kernel function. K is a matrix, called the kernel matrix, whose ijth element is k(xi,xj), where j is an integer in the range 1 to n. N(b, σ ² ) is a probability density function of a Gaussian distribution with mean b and variance σ ² .

The robot control device 2 further includes threshold calculation means 8 which receive information representing the motor position. The torque estimation error model training means 5 outputs information representing the torque estimation error model to the threshold calculation means 8 . The threshold calculation means 8 calculates a threshold based on the torque estimation error model trained by the torque estimation error model training means 5 . Specifically, the threshold calculation means 8 performs the same calculation as the calculation performed by the operating condition calculation means 4 . When the operating condition calculation means 4 has performed a calculation of the engine speed, the threshold value calculation means 8 performs a calculation of the engine speed. When the operating state calculating means 4 has performed an engine acceleration calculation, the threshold value calculating means 8 performs an engine acceleration calculation. When the operating condition calculation means 4 performs a calculation of a part of the components of the driving torque, e.g. the inertial force M(q)a, the threshold calculation means 8 performs a calculation of the inertial force M(q)a. In the first embodiment, the threshold calculation means 8 differentiates the motor position to calculate the motor speed and calculates the speed along each axis of the robot which is equivalent to the motor speed and is represented in the form of an output speed of the gear mechanism along each axis of the robot. The threshold calculation means 8 performs the calculation using equation (2). of the velocity along each axis as x* of equation (2).

The robot control device 2 further includes collision determination means 9. After calculating the equation (2), the threshold value calculation means 8 determines an upper limit value and a lower limit value of a range of ±2σ as threshold values and outputs information representing the threshold values to the collision determination means 9. In calculating the equation (2), the threshold value calculating means 8 may perform a calculation as determined or use a method for reducing the amount of calculation of the Gaussian process regression such as an inducing variable method. The collision determination means 9 determines whether a collision has occurred between the robot and an object by comparing a difference with each of the threshold values calculated by the threshold value calculation means 8, the difference between the drive torque calculated based on a motor current for driving the robot and that of the drive torque calculation means 3 calculated estimated value.

The driving torque calculation means 3 outputs information representing the calculated estimated torque to the collision determination means 9 . The conversion means 6 outputs the information representing the measured torque to the collision determination means 9 . The collision determining means 9 receives the information output from the driving torque calculating means 3, from the converting means 6 and from the threshold calculating means 8. The collision determination means 9 determines that the robot will collide or has collided with an object, and stops the robot when the difference between the measured torque obtained by the converting means 6 and the estimated torque calculated by the drive torque calculation means 3 is greater than or equal to the upper limit value or is less than or equal to the lower limit.

The robot collision detection device 1 according to the first embodiment makes a determination on the occurrence of a collision of a robot with an object by comparing a difference with each of the threshold values calculated by the threshold value calculation means 8, the difference between the drive torque calculated based on a motor current for driving the robot and the estimated value calculated by the drive torque calculation means 3. The threshold calculation means 8 calculates the threshold values based on the torque estimation error model trained by the torque estimation error model training means 5 . This allows the robot collision detection device 1 to set a threshold value considering an effect caused by the torque estimation error not modeled by the driving torque computing means 3, such as the elasticity or non-linearity of the friction of the gear mechanism, and also considering the Variation of the torque estimation error. Thus, when the robot collides with an object, the robot collision detection device 1 can determine the collision of the robot with the object with relatively high sensitivity while preventing misdetection.

Furthermore, the robot collision detection device 1 includes the operational state calculation means 4 which calculates a state quantity related to the operational state of the robot; the torque estimation error model training means 5 which performs the learning using the state quantity calculated by the operating state calculation means 4 as an input signal for the correction function; and the threshold value calculation means 8 which calculates a threshold value based on the torque estimation error model trained by the torque estimation error model training means 5 . This allows the robot collision detection device 1 to set a threshold value considering an effect caused by the torque estimation error correlated with the state quantity related to the operating state of the robot and not modeled by the drive torque calculation means 3, and also under Considering the variation of the torque estimation error. Thus, when the robot collides with an object near the robot, the robot collision detection device 1 can detect the collision of the robot with the object with relatively high sensitivity while taking into account the effect of a factor that has not been modeled and at the same time preventing misdetection .

It is to be noted that in the first embodiment, the operating state calculation means 4 and the threshold value calculation means 8 calculate the speed along each axis, which is equivalent to the engine speed and is represented in terms of the output speed of the gear mechanism. The robot collision detection device 1 may be calculated in the form of the output speed of the gear mechanism instead of the speed along each axis , use the engine speed as the values x1, x2, ..., xn and x*. The robot collision detection apparatus 1 may use the norm of the motor speed along each axis or the norm of the speed along each axis which is equivalent to the motor speed and is represented in terms of the output speed of the gear mechanism as the values x1, x2, . .., use xn and x*.

Second embodiment.

2 12 is a diagram illustrating a configuration of a robot collision detection device 1A according to a second embodiment. The robot collision detection device 1A includes all the components included in the robot collision detection device 1 according to the first embodiment except the conversion means 6. The robot collision detection device 1A includes a robot control device 2A, which includes the driving torque calculation means 3, the threshold value calculation means 8 and the collision determination means 9, but does not include the conversion means 6. The second embodiment will be described with a focus on differences from the first embodiment.

In the second embodiment, the measured torque, which is a measured value of driving torque along each axis of the robot, is measured by a torque sensor provided on a driving unit of the robot. The drive unit and torque sensor are in 2 not illustrated. Information representing the measured torque is received by the error calculation means 7 and the collision determination means 9 . The drive torque calculation means 3 calculates an estimated value of the drive torque without the friction torque of the drive unit or the acceleration/deceleration torque of this motor based on the following equation (5). $$\mathrm{\tau =}{\text{M}}_{\text{L}}\left(\text{q}\right)\text{a}+\text{H}\left(\text{q}\text{, v}\right)+\text{G}\left(\text{q}\right)$$

In Equation (5), M _L (q) is the inertia matrix of the robot excluding the inertia along each axis of the portion closer to the motor than the torque sensor. The above inertia includes the inertia of the motor itself.

The torque estimation error model training means 5 learns a difference and a variation of the difference, which is the difference between the driving torque measured by the torque sensor and an estimated value of the driving torque calculated by the driving torque calculation means 3 . The collision determination means 9 makes a determination about the occurrence of a collision of the robot with an object by comparing the difference with each of the threshold values calculated by the threshold value calculation means 8, the difference between the driving torque measured by the torque sensor and the estimated value calculated by the driving torque calculation means 3 .

The robot collision detection device 1A according to the second embodiment can set a threshold value considering an effect of a factor of the torque estimation error that has not been modeled by the driving torque calculation means 3, such as the elasticity of the gear mechanism and the tension of a cable connected to the robot, and also considering the variation of the torque estimation error, based on the measured torque obtained from the torque sensor, which is not influenced by the engine and by the transmission mechanism. Thus, when the robot collides with an object, the robot collision detection device 1A can determine the collision of the robot with the object with relatively high sensitivity while taking into account the effect of a factor that has not been modeled while preventing misdetection.

Third embodiment.

1 12 is also a diagram illustrating a configuration of the robot collision detection device according to a third embodiment. In the first embodiment, the operating state calculation means 4 receives information representing the motor position, differentiates the motor position to calculate the motor speed, and then calculates, as an operating state quantity, the speed along each axis which is equivalent to the motor speed and is represented in terms of the output speed of the gear mechanism is. In the third embodiment, the operating state calculation means 4 receives the information representing the engine position, performs second-order differentiation of the engine position to calculate the engine acceleration, and then calculates, as an operating state quantity, the acceleration along each axis that is equivalent to the engine acceleration and in Shape of the output acceleration of the gear mechanism is shown.

After receiving the information representing the motor position, the thresholds drive value calculating means 8 performs the same calculation as the calculation performed by the operating condition calculating means 4. Specifically, the threshold calculation means 8 performs second-order differentiation of the motor position to calculate the motor acceleration, and calculates the acceleration along each axis which is equivalent to the motor acceleration and is represented in terms of the output acceleration of the transmission mechanism. Next, the threshold calculation means 8 performs the calculation of Equation (2) using the acceleration along each axis derived for each axis as x* of Equation (2). The details other than the above details are the same as those described in the first embodiment, and the description of the details other than the above details is therefore omitted.

The robot collision detection apparatus according to the third embodiment can correct the effect of the torque estimation error along each axis, which correlates with the acceleration along each axis or the acceleration norm, with relatively high accuracy.

It should be noted that the operating state calculation means 4 and the threshold value calculation means 8 calculate the acceleration along each axis, which is equivalent to the motor acceleration and is represented in terms of the output acceleration of the transmission mechanism. The engine acceleration can be used in the calculation of the equation (2) or in the learning performed in the torque estimation error model training means 5 . Instead of the engine acceleration along each axis, the norm of the engine acceleration along each axis or the norm of the acceleration along each axis that is equivalent to the engine acceleration and is represented in terms of the output acceleration of the gear mechanism as the values x1, x2, ..., xn and x* are used.

Fourth embodiment.

1 12 is also a diagram illustrating a configuration of the robot collision detection device according to a fourth embodiment. In the first embodiment, the operating state calculation means 4 receives information representing the motor position, differentiates the motor position to calculate the motor speed, and then calculates, as an operating state quantity, the speed along each axis which is equivalent to the motor speed and is represented in terms of the output speed of the gear mechanism is. In the fourth embodiment, the operating state calculation means 4 receives the information representing the engine position, differentiates the engine position to calculate the engine speed and the engine acceleration, and then calculates the speed along each axis, which is equivalent to the engine speed and is represented in terms of the output speed of the gear mechanism , and calculate the acceleration along each axis, which is equivalent to the motor acceleration and is represented in terms of the output acceleration of the gear mechanism. The operating state calculation means 4 then calculates the operating state quantity τ based on Equation (6) below. The calculated operational state quantity τ is used in the learning performed in the torque estimation error model training means 5 . $$\mathrm{\tau }=\text{M}\left(\text{q}\right)\text{a}+\text{H}\left(\text{q}\text{, v}\right)+\text{G}\left(\text{q}\right)$$

After receiving the information representing the engine position, the threshold calculation means 8 calculates the engine speed and the engine acceleration in a similar manner to the operating state calculation means 4. The threshold calculation means 8 calculates the speed along each axis, which is equivalent to the engine speed and in the form of the output speed of the gear mechanism, and the acceleration along each axis which is equivalent to the motor acceleration and is expressed in terms of the output acceleration of the gear mechanism, and carry out the calculation of equation (2) using the component along each axis of the operating state quantity τ given on the based on Equation (6) was calculated as x* of Equation (2) by. The details other than the above details are the same as those described in the first embodiment, and the description of the details other than the above details is therefore omitted.

The robot collision detection device according to the fourth embodiment can correct the effect of the driving torque along each axis, or the torque estimation error along each axis in correlation with the driving torque, with relatively high accuracy.

It should be noted that the above description assumes that the operating state calculation means 4 uses the sum of the components M(q)a, h(q,v) and g(q) of equation (6) as the operating state variable. Each of the components M(q)a, h(q,v) and g(q) is part of the components of driving torque, and the frictional force f(v) is not included. The sum of M(q)a, h(q,v) and g(q) is therefore also an example of a Part of the drive torque components. However, the operating state calculation means 4 can use one of M(q)a, h(q,v) and g(q) as the operating state variable. This also provides an example of using a portion of the drive torque components. The threshold calculation means 8 can perform the calculation of Equation (6) or use the calculation result from the driving torque calculation means 3 . In addition, instead of the sum of the components M(q)a, h(q,v) and g(q) of Equation (6), an equal amount as the result of the calculation of Equation (1), which also calculates the frictional force f( v) added, used as the operating state variable. Even if an amount equal to the result of the calculation of Equation (1) is used as the operating state quantity, the threshold calculation means 8 can perform the calculation of Equation (1) or use the calculation result from the drive torque calculation means 3 .

Fifth embodiment.

3 12 is a diagram illustrating a configuration of a robot collision detection device 1B according to a fifth embodiment. The robot collision detection device 1B includes all the components included in the robot collision detection device 1 according to the first embodiment except the drive torque calculation means 3. The robot collision detection device 1B includes the drive torque calculation means 3B instead of the drive torque calculation means 3. The robot collision detection device 1B includes a robot control device 2B having the drive torque calculation means 3B, the conversion means 6, the threshold calculation means 8 and the collision determination means 9. The robot collision detection device 1B further includes parameter identifying means 10 arranged outside the robot control device 2B. The parameter identifying means 10 may be arranged inside the robot control device 2B. The fifth embodiment will be described with a focus on differences from the first embodiment.

The parameter identifying means 10 identifies a value of a parameter of the equation of motion used by the driving torque calculating means 3B based on previously measured data. That is, the parameter identifying means 10 identifies all or part of the parameters of the equation of motion of the robot expressed by Equation (1). When all the parameters of the equation of motion of the robot expressed by Equation (1) are to be identified, the parameter identifying means 10 transforms Equation (1) into Equation (7) below using a vector p. The vector p includes the parameters themselves, such as the mass, the position of the center of gravity, and the coefficient of friction, or what is considered a new parameter that is a result of a calculation performed using two or more parameters, such as "mass × Position of Center of Gravity” is obtained. The parameter identification means 10 then calculates the parameter p using a least squares method based on a vector Y _p derived from the position, from the velocity and from the acceleration at each corresponding instant and the drive torque τ at the instant of the derivation of the vector Y _p . The calculated parameter p is used in the calculation performed by the driving torque calculation means 3B using Equation (1). That is, the driving torque calculating means 3B calculates an estimated value of the driving torque using the values of the parameters identified by the parameter identifying means 10. FIG. $$\mathrm{\tau }=\text{M}\left(\text{q}\right)\text{a}+\text{H}\left(\text{q}\text{, v}\right)+\text{G}\left(\text{q}\right)+\text{f}\left(\text{v}\right)={\text{Y}}_{\text{p}}{}^{\text{T}}\text{p}$$

When a part of the parameters of the robot's equation of motion expressed by Equation (1) is to be identified, the parameter identifying means 10 transforms the equation of motion from Equation (1) into Equation (8) below, where M ₀ (q), h ₀ (q,v), g ₀ (q), and f ₀ (v), respectively, represent an inertia matrix, centrifugal force, gravity, and frictional force calculated on the basis of the unidentifiable parameters, since there are known parameter values, and p ₁ represents a vector containing the parameters to be identified. $$\mathrm{\tau }={\text{M}}_0\left(\text{q}\right)\text{a}+{\text{H}}_0\left(\text{q}\text{, v}\right)+{\text{G}}_0\left(\text{q}\right)+{\text{f}}_0\left(\text{v}\right)+{\text{Y}}_{\text{p1}}{}^{\text{T}}{\text{p}}_1$$

τ ₀ is a value defined by equation (9) below. $${\mathrm{\tau }}_0={\text{M}}_0\left(\text{q}\right)\text{a}+{\text{H}}_0\left(\text{q}\text{, v}\right)+{\text{G}}_0\left(\text{q}\right)+{\text{f}}_0\left(\text{v}\right)$$

The parameter identifying means 10 calculates τ ₁ =τ - τ ₀ based on τ ₀ derived from the position, from the speed and from the acceleration at each corresponding time and the driving torque τ at the time of deriving τ ₀ and calculates the parameter P ₁ using a least squares method together with Y _p1 , the is derived from the position, from the velocity and from the acceleration at each corresponding point in time. The calculated parameter p ₁ is used in the calculation performed by the driving torque calculation means 3B using Equation (1). The driving torque calculating means 3B differs from the driving torque calculating means 3 in that the driving torque calculating means 3B uses the calculated parameter p or the calculated parameter p ₁ .

The robot collision detection device 1B according to the fifth embodiment improves the accuracy of the parameter values of a model related to a dynamic feature of the robot, and when the robot collides with an object, it can detect the collision of the robot with the object with a relatively high sensitivity. while preventing false detection. Moreover, when the robot collides with an object, the robot collision detection device 1</b>B can detect the collision of the robot with the object with relatively high sensitivity while preventing misdetection even when the above parameter values are unknown.

Sixth embodiment.

4 12 is a diagram illustrating a configuration of a robot collision detection device 1C according to a sixth embodiment. The robot collision detection device 1C includes all the components included in the robot collision detection device 1 according to the first embodiment except the driving torque calculation means 3. The robot collision detection device 1C includes the driving torque calculation means 3C instead of the driving torque calculation means 3. The robot collision detection device 1C further includes online parameter identification means 11. The robot collision detection device 1C includes a robot control device 2C , which includes the driving torque calculating means 3C, the converting means 6, the threshold calculating means 8, the collision determining means 9 and the online parameter identifying means 11. The sixth embodiment will be described with a focus on differences from the first embodiment.

The online parameter identifying means 11 identifies a value of a parameter of the equation of motion used by the driving torque calculating means 3C based on data during operation of the robot. For example, the online parameter identification means 11 uses an adaptive identification technique to assign a value of a parameter with an unknown value and a value of a parameter with a variable value during the operation of the robot, among the parameters of the equation of motion of equation (1), during the operation of the robot identify. The on-line parameter identifying means 11 identifies all or part of the parameters of the robot's equation of motion expressed by Equation (1).

$$\text{r}\left[\text{k}\right]=\text{r}\left[\text{k}-1\right]+\text{moit*}\left(-\text{k1*r}\left[\text{k}-1\right]+\mathrm{\tau }\left[\text{k}\right]*{\text{Y}}_{\text{p}}\left[\text{k}\right]\right)$$

$$\text{p}\left[\text{k}\right]=\text{p}\left[\text{k}-1\right]-\text{moit*G1}\cdot \left(\text{R}\left[\text{k}\right]\cdot \text{p}\left[\text{k}-1\right]-\text{r}\left[\text{k}\right]\right)$$

k1 ist ein Gewichtungsfaktor zum Einstellen der Identifizierungsgeschwindigkeit und G1 ist eine Verstärkungsmatrix zum Einstellen der Identifizierungsgeschwindigkeit.When all the parameters of the robot's equation of motion expressed by Equation (1) are to be identified, the online parameter identification means 11 performs identification based on Equations (10), (11) and (12) below, where τ[k] and Y _p [k] represent τ and Y _p of equation (7), respectively, in a k-th identification period, p[k] represents the identified value of p in the k-th identification period, and moit represents the identification period. Note that k is an integer ranging from 1 to n. $$\text{R}\left[\text{k}\right]=\text{R}\left[\text{k}-1\right]+\text{moit*}\left(-\text{k1*R}\left[\text{k}-1\right]+{\text{Y}}_{\text{p}}\left[\text{k}\right]{\text{Y}}_{\text{p}}{\left[\text{k}\right]}^{\text{T}}\right)$$

$$\text{right}\left[\text{k}\right]=\text{right}\left[\text{k}-1\right]+\text{moit*}\left(-\text{k1*r}\left[\text{k}-1\right]+\mathrm{\tau }\left[\text{k}\right]*{\text{Y}}_{\text{p}}\left[\text{k}\right]\right)$$

$$\text{p}\left[\text{k}\right]=\text{p}\left[\text{k}-1\right]-\text{moit*G1}\cdot \left(\text{R}\left[\text{k}\right]\cdot \text{p}\left[\text{k}-1\right]-\text{right}\left[\text{k}\right]\right)$$

k1 is a weighting factor for adjusting identification speed, and G1 is a gain matrix for adjusting identification speed.

When a part of the parameters of the robot's equation of motion expressed by the equation (1) is to be identified, the online parameter identifying means 11 uses τ 1 in the equations (10), (11) and ( ₁₂ ) instead of τ and Y _p1 instead of Y _p to identify the parameter p ₁ instead of p. The on-line parameter identification means 11 outputs information representing the identified parameter values to the driving torque calculation means 3C. The driving torque calculating means 3C receives the information outputted from the online parameter identifying means 11 and uses the parameter values represented in the received information in calculating the equation of motion. That is, the driving torque calculating means 3C uses the values identified by the online parameter identifying means 11 to calculate an estimated value of the drive torque. The driving torque calculating means 3C differs from the driving torque calculating means 3 in that the driving torque calculating means 3C uses the values identified by the on-line parameter identifying means 11 to calculate an estimated value of the driving torque. It should be noted that the on-line parameter identifying means 11 outputs information representing sequentially updated parameter values to the driving torque calculating means 3C.

The robot collision detection device 1C according to the sixth embodiment improves the accuracy of the parameter values of a model related to a dynamic feature of the robot, and when the robot collides with an object, it can detect the collision of the robot with the object with a relatively high sensitivity. while misdetection is prevented even when the value of the parameter related to a dynamic feature of the robot varies during the operation of the robot.

Seventh embodiment.

1 12 is also a diagram illustrating a configuration of the robot collision detection device according to a seventh embodiment. In the first embodiment, the operating state calculation means 4 receives the information representing the motor position, differentiates the motor position to calculate the motor speed, and then calculates, as an operating state variable, the speed along each axis which is equivalent to the motor speed and in terms of the output speed of the gear mechanism is shown. In the seventh embodiment, the operating state calculation means 4 receives information representing the engine position, differentiates the engine position to calculate the engine speed and the engine acceleration, and then calculates the speed along each axis, which is equivalent to the engine speed and is represented in terms of the output speed of the gear mechanism. and the acceleration along each axis, which is equivalent to the motor acceleration and is represented in terms of the output acceleration of the gear mechanism. The operating state calculation means 4 calculates a vector including the velocity along each axis and the acceleration along each axis as an operating state variable.

In contrast to the first embodiment, where the input to the kernel function used by the torque estimation error model training means 5 and by the threshold calculation means 8 has elements of a scalar quantity, in the seventh embodiment the corresponding elements are each a vector quantity. The robot collision detection device according to the seventh embodiment can correct the effect of the torque estimation error along each axis, which correlates with both the acceleration and the velocity along each axis, with relatively high accuracy.

The vectorial acceleration may be replaced with a value obtained by multiplying by a weighting factor the acceleration along each axis which is equivalent to the engine acceleration and is represented in terms of the output acceleration of the gear mechanism. The operating state calculation means 4 can calculate a vector as the operating state variable, which contains not only the speed and the acceleration of each axis to be learned, but also the speeds and accelerations along all axes. The operating state calculation means 4 can calculate a vector as the operating state variable, which contains the positions along all axes, the speeds along all axes and the results of the multiplication of the accelerations along all axes with a weighting.

Eighth embodiment.

5 12 is a diagram illustrating a configuration of a robot collision detection device 1D according to an eighth embodiment. The robot collision detection device 1D includes all components included in the robot collision detection device 1 according to the first embodiment except the threshold value calculation means 8. The robot collision detection device 1D includes the threshold value calculation means 8D instead of the threshold value calculation means 8. The robot collision detection device 1D includes a robot control device 2D having the drive torque calculation means 3, the conversion means 6, the threshold calculation means 8D and the collision determination means 9. The robot collision detection device 1D further includes proximity function training means 12 arranged outside the robot control device 2D. The eighth embodiment will be described with a focus on differences from the first embodiment.

The approximate function training means 12 trains an approximate function based on the torque estimation error model trained by the torque estimation error model training means 5 . In particular, the approximate function training means 12 receives Equation (2) for calculating a predicted distribution and data used to derive Equation (2). When additional data is added, the approximate function training means 12 also receives an output from the operation state calculation means 4. The approximate function training means 12 obtains, through learning, an approximate function that outputs an upper limit value and a lower limit value of a range of ±2σ of an estimated value of the torque estimation error using the Function and the parameter included in Equation (2), the data used for learning by the torque estimation error model training means 5, and the newly added data from the operating state calculation means 4. For example, the approximate function is a feedforward neural network or a recurrent neural network Network.

After performing the training, the approximate function training means 12 outputs information representing the trained approximate function to the threshold value calculating means 8D. The threshold calculation means 8D calculates the threshold values using the approximate function trained by the approximate function training means 12 . That is, the threshold value calculation means 8D calculates the threshold values using the approximate function derived by the approximate function training means 12 . Specifically, after receiving the information representing the motor position, the threshold calculation means 8D performs the same calculation as the calculation performed by the operating condition calculation means 4, inputs the calculation result into the approximate function obtained from the approximate function training means 12, and inputs the calculation result of the approximate function to the collision determination means 9 as the higher and lower thresholds. The higher threshold is the above upper limit and the lower threshold is the above lower limit. The details other than the above details are the same as those described in the first embodiment, and the description of the details other than the above details is therefore omitted.

The robot collision detection device 1D according to the eighth embodiment uses an approximation function instead of a torque estimation error model to calculate the threshold values, and thus can reduce the amount of calculation of the threshold values to a relatively small level, thereby making it possible to calculate the threshold values in a relatively short time.

Ninth embodiment.

6 12 is a diagram illustrating a configuration of a robot collision detection device 1E according to a ninth embodiment. The robot collision detection device 1E includes all the components included in the robot collision detection device 1 according to the first embodiment except the torque estimation error model training means 5 and the threshold calculation means 8. The robot collision detection device 1E includes the torque estimation error model training means 5E instead of the torque estimation error model training means 5 and the threshold calculation means 8E instead of the threshold calculation means 8. The robot collision detection device 1E further includes Temperature measuring means 13 which measure the temperature. For example, the temperature measuring means 13 is a temperature sensor attached to an encoder for measuring the angle of the motor of each axis of the robot. The robot collision detection device 1E includes a robot control device 2E which includes the drive torque calculation means 3, the conversion means 6, the threshold value calculation means 8E, the collision determination means 9 and the temperature measurement means 13. The ninth embodiment will be described with a focus on differences from the first embodiment.

When the robot collision detection device 1E performs the learning, the temperature measuring means 13 outputs information representing a measured temperature to the torque estimation error model training means 5E. The torque estimation error model training means 5E performs the learning using the temperature measured by the temperature measuring means 13 . In the first embodiment, the torque estimation error model training means 5 uses the velocity along each axis of the robot as input to the Gaussian process regression when performing the learning. The torque estimation error model training means 5E uses a vector containing the velocity along each axis and the temperature of each axis as input to the Gaussian process regression for each axis of the robot. The torque estimation error model training means 5E differs from that torque estimation error model training means 5 in that the torque estimation error model training means 5E uses a vector including the velocity along each axis and the temperature of each axis as input to the Gaussian process regression for each axis of the robot. The above temperature of each axis may be replaced with a value obtained by multiplying the above temperature of each axis by a weighting factor.

When the robot collision detection device 1E determines whether a collision between the robot and an object has occurred during the operation of the robot, the temperature measurement means 13 outputs information representing the measured temperature to the threshold value calculation means 8E. The threshold value calculation means 8E calculates the threshold values using the temperature measured by the temperature measurement means 13 . In contrast to the first embodiment, where the input to the kernel function used by the threshold calculation means 8 has elements of velocity along each axis, the threshold calculation means 8E uses a vector representing the velocity along each axis and the temperature of each axis axis included. The threshold value calculation means 8E calculates the threshold values using the temperature measured by the temperature measurement means 13 . The threshold calculation means 8E differs from the threshold calculation means 8 in that the threshold calculation means 8E uses the temperature measured by the temperature measurement means 13 to calculate the threshold values. When the torque estimation error model training means 5E uses a value obtained by multiplying the temperature of each axis by a weighting factor to perform learning, the threshold calculation means 8E uses a vector representing the speed along each axis and that obtained by multiplying the temperature of each axis by a weighting factor value received. The details other than the above details are the same as those described in the first embodiment, and the description of the details other than the above details is therefore omitted.

The robot collision detection device 1E according to the ninth embodiment can improve the effect of the torque estimation error along each axis, which correlates with the temperature of each axis, with relatively high accuracy, and when the robot collides with an object, it can thus avoid the collision of the robot with the object with a relatively high sensitivity while preventing false detection.

7 is a diagram illustrating a processor 71 when at least a part of the driving torque calculation means 3, the operating state calculation means 4, the torque estimation error model training means 5, the conversion means 6, the error calculation means 7, the threshold value calculation means 8 and the collision determination means 9 used in the robot collision detection device 1 according to the first Embodiment are included in the processor 71 is implemented. That is, the functionality of at least part of the driving torque calculation means 3, the operating state calculation means 4, the torque estimation error model training means 5, the conversion means 6, the error calculation means 7, the threshold calculation means 8 and the collision determination means 9 can be implemented in the processor 71, which is a memory 72 executes the saved program.

The processor 71 is a central processing unit (CPU), a processing unit, an arithmetic unit, a microprocessor or a digital signal processor (DSP). In 7 memory 72 is also illustrated.

If the functionality of at least part of the driving torque calculation means 3, the operating state calculation means 4, the torque estimation error model training means 5, the conversion means 6, the error calculation means 7, the threshold calculation means 8 and the collision determination means 9 is implemented in the processor 71, the functionality of at least the part in the Processor 71 implemented by software, firmware, or a combination of software and firmware. The software or firmware is described as a program and is stored in memory 72 . The processor 71 reads and executes a program stored in the memory 72, thereby setting the functionality of at least part of the driving torque calculation means 3, the operating state calculation means 4, the torque estimation error model training means 5, the conversion means 6, the error calculation means 7, the threshold calculation means 8 and the collision determination means 9.

If the functionality of at least part of the driving torque calculation means 3, the operating state calculation means 4, the torque estimation error model training means 5, the conversion means 6, the error calculation means 7, the threshold calculation means 8 and the collision determination means 9 is implemented in the processor 71, the robot collision detection device 1 includes the memory 72 for storing a program that causes the processor 71 to perform at least part of the steps performed by the driving torque calculation means 3, the operating condition calculation means 4, the torque estimation error model training means 5, the conversion means 6, the error calculation means 7, the threshold value calculation means 8 and the collision determination means 9 are to be performed. It can also be said that the program stored in the memory 72 causes a computer to perform at least a part of the operation or method performed by the driving torque calculation means 3, the operating condition calculation means 4, the torque estimation error model training means 5, the conversion means 6, the error calculation means 7, the threshold value calculation means 8 and the collision determination means 9 is to be performed.

The memory 72 is, for example, non-volatile or volatile semiconductor memory such as random access memory (RAM), read-only memory (ROM), flash memory, erasable programmable read-only memory only memory - EPROM) or an electrically erasable programmable read-only memory (EEPROM) (registered trade mark); a magnetic disk, a flexible disk, an optical disk, a compact disk, a mini disk, a digital versatile disk (DVD), or the like.

8th Fig. 12 is a diagram illustrating a processing circuit 81 when at least a part of the driving torque calculation means 3, the operating state calculation means 4, the torque estimation error model training means 5, the conversion means 6, the error calculation means 7, the threshold value calculation means 8 and the collision determination means 9, which are used in the robot collision detection device 1 according to the first embodiment are included in the processing circuit 81 is implemented. That is, at least part of the driving torque calculation means 3, the operating state calculation means 4, the torque estimation error model training means 5, the conversion means 6, the error calculation means 7, the threshold calculation means 8 and the collision determination means 9 can be implemented in the processing circuit 81.

The processing circuit 81 is a dedicated hardware element. The processing circuit 81 is, for example, a single circuit, a set of multiple circuits, a programmed processor, a parallel programmed processor, an application specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or a combination thereof.

A part of the driving torque calculation means 3, the operating state calculation means 4, the torque estimation error model training means 5, the conversion means 6, the error calculation means 7, the threshold calculation means 8 and the collision determination means 9 can be implemented in a dedicated hardware element that is separate from a dedicated hardware element of the rest.

Several functionalities of the driving torque calculation means 3, the operating state calculation means 4, the torque estimation error model training means 5, the conversion means 6, the error calculation means 7, the threshold calculation means 8 and the collision determination means 9 can be implemented such that a part of the several functionalities is implemented in software or firmware and the rest of the several functionalities is implemented in a dedicated hardware element. Thus, several functionalities of the driving torque calculation means 3, the operating state calculation means 4, the torque estimation error model training means 5, the conversion means 6, the error calculation means 7, the threshold calculation means 8 and the collision determination means 9 can be implemented in hardware, software, firmware or a combination thereof.

The functionality of at least part of the driving torque calculation means 3, the operating state calculation means 4, the torque estimation error model training means 5, the error calculation means 7, the threshold calculation means 8 and the collision determination means 9 included in the robot collision detection device 1A according to the second embodiment can be implemented in a processor that executes a program stored in memory. Such memory is memory similar to memory 72, and such processor is processor 71-like processor. At least a part of the above-described driving torque calculation means 3, operating state calculation means 4, torque estimation error model training means 5, error calculation means 7, threshold value be Calculation means 8 and the collision determination means 9 can be implemented in a processing circuit. Such a processing circuit is a processing circuit similar to the processing circuit 81 .

The functionality of at least part of the driving torque calculation means 3, the operating state calculation means 4, the torque estimation error model training means 5, the conversion means 6, the error calculation means 7, the threshold value calculation means 8 and the collision determination means 9 included in each of the robot collision detection devices according to the third, fourth and seventh embodiments, may be implemented in a processor executing a program stored in memory. Such memory is memory similar to memory 72, and such processor is processor 71-like processor. At least part of the driving torque calculation means 3, the operating state calculation means 4, the torque estimation error model training means 5, the conversion means 6, the error calculation means 7, the threshold value calculation means 8 and the collision determination means 9 included in each of the robot collision detection devices according to the third, fourth and seventh embodiments may be in a Processing circuit be implemented. Such a processing circuit is a processing circuit similar to the processing circuit 81 .

The functionality of at least a part of the driving torque calculation means 3B, the operating state calculation means 4, the torque estimation error model training means 5, the conversion means 6, the error calculation means 7, the threshold value calculation means 8, the collision determination means 9 and the parameter identification means 10 included in the robot collision detection device 1B according to the fifth embodiment, may be implemented in a processor executing a program stored in memory. Such memory is memory similar to memory 72, and such processor is processor 71-like processor. At least part of the above-described driving torque calculation means 3B, operating state calculation means 4, torque estimation error model training means 5, conversion means 6, error calculation means 7, threshold calculation means 8, collision determination means 9 and parameter identification means 10 may be implemented in a processing circuit. Such a processing circuit is a processing circuit similar to the processing circuit 81 .

The functionality of at least part of the driving torque calculation means 3C, the operating state calculation means 4, the torque estimation error model training means 5, the conversion means 6, the error calculation means 7, the threshold calculation means 8, the collision determination means 9 and the online parameter identification means 11 included in the robot collision detection device 1C according to the sixth embodiment may be implemented in a processor executing a program stored in memory. Such memory is memory similar to memory 72, and such processor is processor 71-like processor. At least part of the above-described drive torque calculation means 3C, operating state calculation means 4, torque estimation error model training means 5, conversion means 6, error calculation means 7, threshold calculation means 8, collision determination means 9 and online parameter identification means 11 may be implemented in a processing circuit. Such a processing circuit is a processing circuit similar to the processing circuit 81 .

The functionality of at least a part of the driving torque calculation means 3, the operating state calculation means 4, the torque estimation error model training means 5, the conversion means 6, the error calculation means 7, the threshold value calculation means 8D, the collision determination means 9 and the approximate function training means 12 included in the robot collision detection device 1D according to the eighth embodiment, may be implemented in a processor executing a program stored in memory. Such memory is memory similar to memory 72, and such processor is processor 71-like processor. At least part of the above-described driving torque calculation means 3, operating state calculation means 4, torque estimation error model training means 5, conversion means 6, error calculation means 7, threshold calculation means 8D, collision determination means 9 and approximate function training means 12 may be implemented in a processing circuit. Such a processing circuit is a processing circuit similar to the processing circuit 81 .

The functionality of at least part of the driving torque calculation means 3, the operating state calculation means 4, the torque estimation error model training means 5E, the conversion means 6, the error calculation means 7, the threshold calculation means 8E, the collision determination means 9 and the temperature measurement means 13 used in the robot collision detection tion device 1E according to the ninth embodiment may be implemented in a processor that executes a program stored in a memory. Such memory is memory similar to memory 72, and such processor is processor 71-like processor. At least part of the above-described driving torque calculation means 3, operating state calculation means 4, torque estimation error model training means 5E, conversion means 6, error calculation means 7, threshold calculation means 8E, collision determination means 9 and temperature measurement means 13 may be implemented in a processing circuit. Such a processing circuit is a processing circuit similar to the processing circuit 81 .

The configurations described in the above embodiments are just examples. These configurations can be combined with other known technology, and configurations of different embodiments can be combined with each other. Moreover, part of the configurations can be omitted and/or modified without departing from the spirit.

Reference List

1, 1A, 1B, 1C, 1D, 1E

robot collision detection device;

2, 2A, 2B, 2C, 2D, 2E

robot control device;

3, 3B, 3C

drive torque calculation means;

4

operating condition calculation means;

5, 5E

torque estimation error model training means;

5a

training aids;

6

conversion agents;

7

error calculation means;

8, 8D, 8E

threshold calculation means;

9

collision determination means;

10

parameter identification means;

11

online parameter identification means;

12

approximate function training aids;

13

temperature measuring means;

71

Processor;

72

Storage;

81

processing circuit.

QUOTES INCLUDED IN DESCRIPTION

This list of documents cited by the applicant was generated automatically and is included solely for the better information of the reader. The list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions.

Patent Literature Cited

• JP 2016511699 [0004]

Claims (12)

A robot collision detection device comprising: driving torque calculation means for calculating an estimated value of the driving torque of a robot; torque estimation error model training means for learning a difference and a variation of the difference, the difference between the drive torque calculated based on a motor current for driving the robot and the estimated value calculated by the drive torque calculation means; threshold calculation means for calculating a threshold value based on a torque estimation error model trained by the torque estimation error model training means; and Collision determination means for making a determination on the occurrence of a collision of the robot with an object by comparing the difference with the threshold value calculated by the threshold calculation means, the difference between the driving torque calculated based on the motor current for driving the robot and the estimated one calculated by the driving torque calculation means value is present. A robot collision detection device comprising: driving torque calculation means for calculating an estimated value of the driving torque of a robot; torque estimation error model training means for learning a difference and a variation of the difference, the difference between the driving torque measured by a torque sensor provided on a driving unit of the robot and the estimated value calculated by the driving torque calculating means; threshold calculation means for calculating a threshold value based on a torque estimation error model trained by the torque estimation error model training means; and Collision determination means for making a determination as to occurrence of a collision of the robot with an object by comparing the difference with the threshold value calculated by the threshold calculation means, the difference between the driving torque measured by the torque sensor and the estimated value calculated by the driving torque calculation means. Robot collision detection device claim 1 or 2 , wherein the torque estimation error model training means includes training means using a non-parametric technique. Robot collision detection device claim 3 , where the training means use Gaussian process regression. Robot collision detection device according to any one of Claims 1 until 4 , further comprising: parameter identifying means for identifying, based on previously measured data, a value of a parameter of an equation of motion to be used by the driving torque calculating means, wherein the driving torque calculating means calculates the estimated value using the value identified by the parameter identifying means. Robot collision detection device according to any one of Claims 1 until 4 , further comprising: online parameter identifying means for identifying, based on data during operation of the robot, a value of a parameter of an equation of motion to be used by the driving torque calculating means, wherein the driving torque calculating means uses the value identified by the online parameter identifying means to calculate the use estimated value. Robot collision detection device according to any one of Claims 1 until 6 further comprising: operation state calculation means for calculating a state quantity related to an operation state of the robot, wherein the torque estimation error model training means performs learning using the state quantity calculated by the operation state calculation means as an input signal for a correction function. Robot collision detection device claim 7 , wherein the operating state calculation means calculates the state quantity including information about an engine speed. Robot collision detection device claim 7 , wherein the operating state calculation means calculates the state quantity including information about an engine acceleration. Robot collision detection device claim 7 , wherein the operating state calculation means the state variable, which is a part of Components of the drive torque of the robot includes calculate. Robot collision detection device according to any one of Claims 1 until 10 , further comprising: approximate function training means for training an approximate function based on the torque estimation error model trained by the torque estimation error model training means, wherein the threshold value calculation means calculates the threshold value using the approximate function derived by the approximate function training means. Robot collision detection device according to any one of Claims 1 until 11 , further comprising: temperature measurement means for measuring a temperature, wherein the torque estimation error model training means performs learning using the temperature measured by the temperature measurement means, and the threshold value calculation means calculates the threshold value using the temperature measured by the temperature measurement means.

Images