Distributional expert demonstrations for autonomous driving
The present disclosure provides methods and techniques for evaluating and improving algorithms for autonomous driving planning and control (PNC), using one or more metrics (e.g., similarity scores) computed based on expert demonstrations. For example, the one or more metrics allow for improving PNC based on human, as opposed to or in addition to optimizing certain oversimplified properties, such as the least distance or time, as an objective. When driving in certain scenarios, such as taking a turn, people may drive in a distributed probability pattern instead of in a uniform line (e.g., different speeds and different curvatures at the same corner). As such, there can be more than one “correct” control trajectory for an autonomous vehicle to perform in the same turn. Safety, comfort, speeds, and other criteria may lead to different preferences and judgment as to how well the controlled trajectory has been computed.
1. A computer-implemented method comprising:
generating, by a processing device, a controlled trajectory of an autonomous driving vehicle (ADV) in a scenario, the controlled trajectory executable by the ADV to drive autonomously in the scenario;
receiving a set of data acquired in a plurality of driving demonstrations in the scenario;
identifying a distribution pattern of the set of data acquired, the distribution pattern indicating probabilities of driving trajectories in the scenario;
computing, by the processing device, a similarity score based on comparisons between the controlled trajectory and the distribution pattern of the plurality of driving demonstrations;
computing a first mean square error and a first similarity score between a first controlled trajectory generated by the processing device and the set of data of the plurality of driving demonstrations in the scenario;
computing a second mean square error and a second similarity score between a second controlled trajectory generated by the processing device and the set of data of the plurality of driving demonstrations in the scenario, wherein the first mean square error is greater than the second mean square error while the first similarity score is higher than the second similarity score; and
controlling, by the processing device, the ADV based on the first and the second mean square errors and the first and the second similarity scores.
2. The computer-implemented method of claim 1 , further comprising:
adjusting, based on the similarity score, at least one parameter in generating the controlled trajectory of the ADV in the scenario; and
generating an updated controlled trajectory, wherein the updated controlled trajectory corresponds to an updated similarity score higher than the similarity score computed previously, the updated controlled trajectory executable by the ADV to drive autonomously in the scenario.
3. The computer-implemented method of claim 1 , wherein generating the controlled trajectory of the ADV in the scenario comprises:
generating a plurality of candidate trajectories subject to a probability distribution pattern.
4. The computer-implemented method of claim 3 , wherein computing the similarity score comprises at least one of:
calculating a Kullback-Leibler divergence;
calculating an f-divergence;
calculating an H-divergence;
determining one or more integral probability metrics; or
evaluating a probability density of a predicted value of a probability distribution function of the distribution pattern.
5. The computer-implemented method of claim 1 , wherein the scenario comprises:
a starting location and an initial speed vector;
an ending location and a final speed vector; and
at least one of: a static side boundary, an available width along the static side boundary, a dynamic side boundary, or a width variation caused by the dynamic side boundary.
6. The computer-implemented method of claim 5 , wherein the scenario comprises at least one of: a turn or a sequence change of directions, an overtake, a compliance to traffic signs, or a timed attempt thereof.
7. The computer-implemented method of claim 1 , wherein the distribution pattern comprises an observed feature distribution of distances, speeds, or accelerations, or an embedding feature distribution of latent space features in a deep learning layer that processes the set of data acquired in the plurality of driving demonstrations.
8. The computer-implemented method of claim 1 , wherein the plurality of driving demonstrations comprises demonstrations performed in a simulation environment.
9. A non-transitory machine-readable medium having instructions stored therein, which when executed by a processor, cause the processor to perform operations, the operations comprising:
generating a controlled trajectory of an autonomous driving vehicle (ADV) in a scenario, the controlled trajectory executable by the ADV to drive autonomously in the scenario;
receiving a set of data acquired in a plurality of driving demonstrations in the scenario;
identifying a distribution pattern of the set of data acquired, the distribution pattern indicating probabilities of driving trajectories in the scenario;
computing a similarity score based on comparisons between the controlled trajectory and the distribution pattern of the plurality of driving demonstrations;
computing a first mean square error and a first similarity score between a first controlled trajectory generated by the processor and the set of data of the plurality of driving demonstrations in the scenario;
computing a second mean square error and a second similarity score between a second controlled trajectory generated by the processor and the set of data of the plurality of driving demonstrations in the scenario, wherein the first mean square error is greater than the second mean square error while the first similarity score is higher than the second similarity score; and
controlling, by the processor, the ADV based on the first and the second mean square errors and the first and the second similarity scores.
10. The machine-readable medium of claim 9 , wherein the operations further comprise:
adjusting, based on the similarity score, at least one parameter in generating the controlled trajectory of the ADV in the scenario; and
generating an updated controlled trajectory, wherein the updated controlled trajectory corresponds to an updated similarity score higher than the similarity score computed previously, the updated controlled trajectory executable by the ADV to drive autonomously in the scenario.
11. The machine-readable medium of claim 9 , wherein generating the controlled trajectory of the ADV in the scenario comprises:
generating a plurality of candidate trajectories subject to a probability distribution pattern.
12. The machine-readable medium of claim 11 , wherein computing the similarity score comprises at least one of:
calculating a Kullback-Leibler divergence;
calculating an f-divergence;
calculating an H-divergence;
determining one or more integral probability metrics; or
evaluating a probability density of a predicted value of a probability distribution function of the distribution pattern.
13. The machine-readable medium of claim 9 , wherein the scenario comprises:
a starting location and an initial speed vector;
an ending location and a final speed vector; and
at least one of: a static side boundary, an available width along the static side boundary, a dynamic side boundary, or a width variation caused by the dynamic side boundary.
14. The machine-readable medium of claim 13 , wherein the scenario comprises at least one of: a turn or a sequence change of directions, an overtake, a compliance to traffic signs, or a timed attempt thereof.
15. The machine-readable medium of claim 9 , wherein the distribution pattern comprises an observed feature distribution of distances, speeds, or accelerations, or an embedding feature distribution of latent space features in a deep learning layer that processes the set of data acquired in the plurality of driving demonstrations.
16. The machine-readable medium of claim 9 , wherein the plurality of driving demonstrations comprises demonstrations performed in a simulation environment.
17. A processing device, comprising:
a processor; and
a memory coupled to the processor and storing instructions, which when executed by the processor, cause the processor to
generate a controlled trajectory of an autonomous driving vehicle (ADV) in a scenario, the controlled trajectory executable by the ADV to drive autonomously in the scenario;
receive a set of data acquired in a plurality of driving demonstrations in the scenario;
identify a distribution pattern of the set of data acquired, the distribution pattern indicating probabilities of driving trajectories in the scenario;
compute, by the processing device, a similarity score based on comparisons between the controlled trajectory and the distribution pattern of the plurality of driving demonstrations;
compute a first mean square error and a first similarity score between a first controlled trajectory generated by the processing device and the set of data of the plurality of driving demonstrations in the scenario;
compute a second mean square error and a second similarity score between a second controlled trajectory generated by the processing device and the set of data of the plurality of driving demonstrations in the scenario, wherein the first mean square error is greater than the second mean square error while the first similarity score is higher than the second similarity score; and
control, by the processing device, the ADV based on the first and the second mean square errors and the first and the second similarity scores.
18. The processing device of claim 17 , wherein the instructions further cause the processor to:
adjust, based on the similarity score, at least one parameter in generating the controlled trajectory of the ADV in the scenario; and
generate an updated controlled trajectory, wherein the updated controlled trajectory corresponds to an updated similarity score higher than the similarity score computed previously, the updated controlled trajectory executable by the ADV to drive autonomously in the scenario.