Systems and methods for vehicles navigating roads using a control model trained with residual policies

A non-transitory computer-readable medium comprising:

• instructions that when executed by a processor cause the processor to:
• generate a generic policy for a control model to navigate a road having multiple agents with traffic data, the generic policy applying to general traffic scenarios associated with the road;
• train a task policy by a reinforcement learning model learning residual functions of the generic policy using a neural network, the residual functions factoring parameters about the multiple agents and specific traffic scenarios and the task policy includes a decision task with segmented actions on the road having different tasks for the generic policy and the residual functions;
• learn an intermediate policy for the generic policy used by the control model associated with an unmet stopping distance on the road between the multiple agents;
• and
• communicate the generic policy, the task policy for error reduction, and a domain distribution learned for a lane geometry associated with the road by comparing simulated data with the traffic data to a vehicle.

A method comprising:

• generating a generic policy for a control model to navigate a road having multiple agents with traffic data, the generic policy applying to general traffic scenarios associated with the road;
• training a task policy by a reinforcement learning model learning residual functions of the generic policy using a neural network, the residual functions factoring parameters about the multiple agents and specific traffic scenarios and the task policy includes a decision task with segmented actions on the road having different tasks for the generic policy and the residual functions;
• learning an intermediate policy for the generic policy used by the control model associated with an unmet stopping distance on the road between the multiple agents;
• and
• communicating the generic policy, the task policy for error reduction, and a domain distribution learned for a lane geometry associated with the road by comparing simulated data with the traffic data to a vehicle.