DE102021201011A1 - Unsupervised optimization of odometry models - Google Patents

Abstract

The invention relates to a method for calculating a pose of a vehicle; wherein the pose or change in pose of the vehicle is measured by one or more sensors; wherein the pose or change in pose of the vehicle is calculated using a mechanical model of the vehicle; and wherein a deviation of the calculated pose from the measured pose is determined. The model is optimized through unsupervised reinforcement learning, with deviation as a reward.

Description

The invention relates to a method according to the preamble of claim 1, a computer program according to claim 4 and a computing device according to claim 5.

For the optimal localization and navigation of autonomous vehicles, a high accuracy of the prediction of a vehicle's own movement (Vehicle Motion Estimation, VME) is necessary. The VME is often realized by fusing the data from different sensors, e.g. with the help of Kalman filters. The data is processed using expert models.

A large number of algorithms or methods (e.g. SLAM) exist for localization and navigation. The calculation of the current position of the vehicle in a defined space depends on the quality of the environment sensing. The quality of the targets measured by a sensor can often only be estimated with difficulty. It is usually unclear which sensor should be used under which environmental conditions, i.e. which sensor can be trusted.

To evaluate the sensor quality in the current period, static values are assumed and defined in a so-called covariance matrix. In order to merge the data from different sensors, precise knowledge of the covariances, i.e. the mutual dependencies between the sensors and algorithms used, is necessary. However, these are usually unknown or can only be determined very imprecisely.

If a vehicle frequently moves in a similar state space, knowledge gained about the environment has not been permanently recorded so far. As a result, similar situations and conditions have to be reassessed every time you drive.

The use of neural networks to solve the forward kinematics of a 3-DOF actuator is known from the prior art. A corresponding approach is described in "Kinematic analysis of a novel 3-DOF actuation redundant parallel manipulator using artificial intelligence approach" (Zhang D., Lei J.; Robotics and Computer-Integrated Manufacturing, 27(1), 2001, 157-163 ).

In "Design optimization of a spatial six degree-of-freedom parallel manipulator based on artificial intelligence approaches" (Gao Z., Zhang D., Ge Y; Robotics and Computer-Integrated Manufacturing, 26(2), 2010, 180-189 ) the use of artificial intelligence is described to describe the rigidity and flexibility of a 6-DOF robot arm and the kinematics. This optimizes the design of the actuator.

A systematic method for manual correction and optimization of odometry of mobile robots is from "Correction of systematic odometry errors in mobile robots" (Borenstein, J., Feng, L.; in Proceedings 1995 IEEEIRSJ International Conference on Intelligent Robots and Systems. Human Robot Interaction and Cooperative Robots, Vol. 3, 569-574, IEEE, 1995).

The pamphlet DE 10 2018 002 029 A1 refers to determining a prediction for a pose of a vehicle. Measurement data from the past are recorded and learned using a neural network. The neural network uses the learned measurement data to determine a future pose for the vehicle.

The object of the invention is to improve the accuracy of determining a pose of a vehicle on the basis of a mechanical model. This object is achieved by a method according to claim 1, a computer program according to claim 4 and a computing device according to claim 5. Preferred developments are contained in the dependent claims and result from the following description.

The method according to the invention serves to calculate a pose of a vehicle. The pose is understood to mean a position and orientation of the vehicle.

The pose or a change in pose is first measured by one or more sensors. NSS, lidar and/or radar sensors are used as sensors, for example. The sensors and their measurements form a reference system.

In addition to the measured pose, the pose or change in pose is calculated using a mechanical model of the vehicle. The mechanical model is a model that describes one or more mechanical properties of the vehicle. The mechanical properties include the kinematic properties of the vehicle, which relate to movements of the vehicle without the influence of forces, and the dynamic properties of the vehicle, which reflect the influence of forces acting on or within the vehicle. In particular, the mechanical model can be designed as a kinematic model that only takes into account the kinematics of the vehicle.

According to the invention, the model is optimized by unsupervised reinforcement learning. A previously determined deviation of the calculated pose from the measured pose serves as a reward.

In the event of a failure of the reference system, the model is used to determine the pose of the vehicle. The invention makes it possible to react very well to new circumstances in the environment and to include time profiles. Complex odometry models can be created more easily in this way and depict reality more precisely thanks to the learning method according to the invention. With the invention, an odometry model can be dynamically adapted to changes in the loading condition of the vehicle and a resulting shift in the center of gravity, for example.

The mechanical model of the vehicle can be designed as a black box or gray box model. If a gray box model is used, the artificial intelligence learns an optimal configuration of the model and a dynamic variation of the parameters. It is based on a deterministic behavior. In this case, the model itself is known and is not subject to any learning process.

In a black box model, the artificial intelligence learns the mechanical model entirely by itself. For this purpose, randomized initial weights or pre-adjusted weights are used to roughly estimate a known model and lead to faster convergence. The learned model is unknown in this case. Restrictions in the form of a given model can be avoided by using a black box model.

Using the invention, a mechanical model for a vehicle fleet can be trained, for example, by first subjecting the model to preparatory training with a limited number of parameters until a defined accuracy is achieved. A first neural network is preferably trained as part of the preparatory training. A second neural network is then used. The first neural network can be used on all vehicles in the fleet. The weights are optimized via a cloud if a reference system is available. However, the weight changes made are now very small. The output of the first neural network is adjusted by means of a further, vehicle-specific neural network, taking into account further parameters, in order to adapt it to vehicle-specific properties, for example different locations, varying equipment or different vehicle weight.

In a preferred development, the method is carried out iteratively. As iteration progresses, the model always adapts precisely to the vehicle's conditions. A failure of the reference system can thus be compensated better and over a longer period of time.

In another preferred development, at least one value of a physical variable of the vehicle is measured and used as an input parameter for the mechanical model. Possible physical parameters include odometry data from the vehicle.

The computer program according to the invention is designed to cause a computing device to carry out the method according to the invention or a preferred development.

A computing device according to the invention is designed to carry out the method according to the invention or a preferred development. A computing device is in accordance with the invention, for example, if it contains a computer program in accordance with the invention.

By optimizing the model using unsupervised reinforcement learning, an assessment of the detection probabilities of the individual sensors in the current period can be calculated. In a preferred development, a detection probability of at least one sensor is calculated accordingly. The quality assessment of the relative localization takes place in real time and is assigned to a position and stored on a digital map. This makes it possible to detect whether the sensors deliver plausible values. The quality assessment of the localization is stored in a localization quality map for the current position. Over time, a data structure is created from which the sensor quality can be read out depending on the vehicle condition.

In a preferred exemplary embodiment of the invention, the change in a position of a vehicle as a function of wheel speeds and defined influencing parameters is predicted using an artificial intelligence method. During the training, the localization is corrected by a reference system, for example GPS. The quality of the reference system is also taken into account.

und initialem Score wie folgt trainiert:

• - Durch Einbeziehung von Stochastiken werden die „erlernten“ Gewichtungen beeinflusst. Daraus ergibt sich ein evolutionäres Verhalten. Mit jeder Epoche werden mehrere Modelle parallel trainiert. Als Ausgangselement für die nächste Epoche werden dabei jeweils die besten Kandidaten selektiert.
• - In jedem Schritt t wird mittels künstlicher Intelligenz eine Positionsänderung $\mathrm{\Delta }{p}_{t+1}^t$
  
  in Abhängigkeit von s_t vorhergesagt. $$p_{t+1}^s=p_t^s+\mathrm{\Delta }{p}_{t+1}^t\left(s_t\right)$$
  
  Mittels einer Feedback-Funktion wird eine Belohnung $r_t\left(p_{t+1}^s,p_{t+1}\right)$
  
  generiert. Dabei erfolgt jeweils eine Backpropagation zur Justierung der Gewichtungen der Verbindungen der „Neuronen“ der künstlichen Intelligenz.
• - Die jeweilige Epoche wird beendet, sobald eine maximale Abweichung von $dist\left(p_{t+1}^s,p_{t+1}\right)$
  
  überschritten wird. Hiernach ist die maximal fahrbare Distanz bekannt, die im Schnitt zuverlässig gefahren werden kann (Qualitätsbeschreibung).
• - Durch Summation ergibt sich eine Gesamtbelohnung $r={\displaystyle {\sum }_{t=0}^{t_{max}}{r}_t}.$
  
  Die besten Kandidaten werden ausgewählt.
• - Als Feedback wird z.B. folgender Wert einer Glockenkurve verwendet: $$\frac{1}{\#Parameter::Genauigkeit}\cdot \text{exp}\left(-\#Parameter::Genauigkei{t}^2\cdot dist{\left(p_{t+1}^s,p_{t+1}\right)}^2\right)$$
  

If the quality of the reference system is sufficient, the method starts from an initial start state s ₀ with a known position $p_0^s=p_0$

and initial score trained as follows:

• - The "learned" weightings are influenced by the inclusion of stochastics. This results in an evolutionary behavior. With each epoch, multiple models are trained in parallel. The best candidates are selected as the starting point for the next epoch.
• - In each step t, a position change is made using artificial intelligence $\mathrm{\Delta }{p}_{t+1}^t$
  
  predicted as a function of s _t . $$p_{t+1}^s=p_t^s+\mathrm{\Delta }{p}_{t+1}^t\left(s_t\right)$$
  
  A reward is provided by means of a feedback function ${\mathrm{right}}_t\left(p_{t+1}^s,p_{t+1}\right)$
  
  generated. In each case, a back propagation takes place to adjust the weighting of the connections of the "neurons" of the artificial intelligence.
• - The respective epoch ends as soon as a maximum deviation of $\mathrm{i.e}ist\left(p_{t+1}^s,p_{t+1}\right)$
  
  is exceeded. The maximum distance that can be driven reliably on average is then known (quality description).
• - Summation results in a total reward $\mathrm{right}={\displaystyle {\sum }_{t=0}^{t_{max}}{\mathrm{right}}_t}.$
  
  The best candidates are selected.
• - The following value of a bell curve is used as feedback: $$\frac{1}{\#Pa\mathrm{right}amete\mathrm{right}::Gena\mathrm{and}iGkeit}\cdot \text{ex}\left(-\#Pa\mathrm{right}amete\mathrm{right}::Gena\mathrm{and}iGkei{t}^2\cdot \mathrm{i.e}ist{\left(p_{t+1}^s,p_{t+1}\right)}^2\right)$$
  

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

• DE 102018002029 A1 [0009]

Claims (7)

Method for calculating a pose of a vehicle; wherein the pose or change in pose of the vehicle is measured by one or more sensors; wherein the pose or change in pose of the vehicle is calculated using a mechanical model of the vehicle; and wherein a deviation of the calculated pose from the measured pose is determined; characterized in that the model is optimized by unsupervised reinforcement learning, with deviation as a reward. procedure after claim 1 ; characterized in that the method is carried out iteratively. Method according to one of the preceding claims; characterized in that at least one value of a physical quantity of the vehicle is measured and used as an input parameter of the mechanical model. Method according to one of the preceding claims; characterized in that at least one value of a physical quantity of the vehicle environment is measured and sent as an input parameter of the mechanical model. Method according to one of the preceding claims; characterized in that a detection probability of at least one sensor is calculated. Computer program designed to cause a computing device to execute a method according to one of the preceding claims. Computing device which is designed to carry out a method according to one of the preceding method claims.