Vehicle adaptive learning
A computing system can determine a vehicle action based on inputting vehicle sensor data to a first neural network including a first safety agent that can determine a probability of unsafe vehicle operation. The first neural network can be adapted, at a plurality of times, by a periodically retrained deep reinforcement learning agent that includes a second deep neural network including a second safety agent. A vehicle can be operated based on the vehicle action.
1. A method, comprising:
inputting vehicle sensor data to a first neural network (NN) that outputs vehicle transition states that include a predicted 3D pose, speed, and lateral and longitudinal acceleration data;
wherein the NN includes a safety agent that based on the transition states, determines a probability of unsafe vehicle operation and is adapted, at a plurality of times, by a periodically retrained deep reinforcement learning agent that includes a second NN including a second safety agent,
wherein the second NN and second safety agent receive as input at least some of the transition states, information about safety violations that include a collision or near collision, and one or more termination states that have been substituted for one or more of the transition states, and
wherein a termination state includes a vehicle operation to avoid safety violations; and
operating a vehicle based on a vehicle action output from the first NN.
2. The method of claim 1 , wherein the vehicle action includes operation of vehicle steering, braking, and powertrain components.
3. The method of claim 1 , further comprising inputting vehicle sensor data by inputting a color video image into the first NN.
4. The method of claim 1 , wherein the safety agent determines probabilities of unsafe vehicle operation based on inputting vehicle actions into one of a rule-based machine learning system or a third NN, trained based on simulated data.
5. The method of claim 1 , wherein the second NN is trained based on vehicle action ground truth, wherein vehicle action ground truth includes vehicle sensor data, vehicle actions, and information regarding unsafe vehicle operation from the safety agent.
6. The method of claim 5 , wherein the vehicle sensor data, vehicle actions, and information from the safety agent is based on simulated data.
7. The method of claim 1 , further comprising periodically retraining the second NN, a reinforcement learning agent, based on recorded vehicle sensor data, recorded vehicle action ground truth, and recorded information regarding unsafe vehicle operation from the safety agent, subject to an error bound.
8. The method of claim 7 , further comprising adapting, at a plurality of times, the first NN by updating first NN parameters with parameters from the second NN.
9. The method of claim 1 , where the initial training of first NN is based on both safe buffer (state, safe action) pairs and unsafe buffer (state, unsafe action) pairs collected during offline simulation and training of the second NN is based on both safe buffer (state, safe action) pairs and unsafe buffer (state, unsafe action) collected during offline simulation and collected during driving.
10. A system, comprising a processor; and
a memory, the memory including instructions to be executed by the processor to:
input vehicle sensor data to a first neural network (NN) that outputs vehicle transition states that include a predicted 3D pose, speed, and lateral and longitudinal acceleration data;
wherein the NN includes a safety agent that based on the transition states, determines a probability of unsafe vehicle operation and is adapted, at a plurality of times, by a periodically retrained deep reinforcement learning agent that includes a second NN including a second safety agent,
wherein the second NN and second safety agent receive as input at least some of the transition states, information about safety violations that include a collision or near collision, and one or more termination states that have been substituted for one or more of the transition states, and
wherein a termination state includes a vehicle operation to avoid safety violations; and
and
operate a vehicle based on the vehicle action output from the first NN.
11. The system of claim 10 , wherein the vehicle action includes operation of vehicle steering, braking, and powertrain components.
12. The system of claim 10 , further comprising inputting vehicle sensor data by inputting a color video image into the first deep neural network.
13. The system of claim 10 , wherein the safety agent determines probabilities of unsafe vehicle operations based on inputting vehicle actions into one of a rule-based machine learning system or a third NN, trained based on simulated data.
14. The system of claim 10 , wherein the second NN is trained based on vehicle action ground truth, wherein vehicle action ground truth includes vehicle sensor data, vehicle actions, and information regarding unsafe vehicle operation from the safety agent.
15. The system of claim 14 , wherein the vehicle sensor data, vehicle actions, and information regarding unsafe vehicle operations from the safety agent is based on simulated data.
16. The system of claim 10 , further comprising periodically retraining the second NN, deep reinforcement learning agent, based on recorded vehicle sensor data, recorded vehicle action ground truth, and recorded information regarding unsafe vehicle operation from the safety agent, subject to an error bound.
17. The system of claim 16 , further comprising adapting, at a plurality of times, the first NN by updating first NN parameters with parameters from the second NN.
18. The system of claim 10 where the initial training of first NN is based on both safe buffer (state, safe action) pairs and unsafe buffer (state, unsafe action) pairs collected during offline simulation and training of the second NN is based on both safe buffer (state, safe action) pairs and unsafe buffer (state, unsafe action) collected during offline simulation and collected during driving.
19. A system, comprising:
means for controlling vehicle steering, braking and powertrain; and
means for:
inputting vehicle sensor data to a first neural network (NN) including a first safety agent that outputs vehicle transition states that include a predicted 3D pose, speed, and lateral and longitudinal acceleration data;
wherein the NN includes a safety agent that based on the transition states, determines a probability of unsafe vehicle operation and is adapted, at a plurality of times, by a periodically retrained deep reinforcement learning agent that includes a second NN including a second safety agent,
wherein the second NN and second safety agent receive as input at least some of the transition states, information about safety violations that include a collision or near collision, and one or more termination states that have been substituted for one or more of the transition states, and
wherein a termination state includes a vehicle operation to avoid safety violations; and
operating a vehicle based on the vehicle action and means for controlling vehicle steering, braking and powertrain.
20. The system of claim 19 , further comprising inputting vehicle sensor data by inputting a color video image into the first NN.