Abstract: A method for generating a motion plan for an autonomous agent includes obtaining scene data for an agent's environment, including map data and historical state data for one or more objects. A transformer-based encoder generates a set of scene embedding tokens from the scene data to represent a fixed environmental context. A transformer-based decoder autoregressively generates a sequence of action tokens representing a future trajectory. The generation of each subsequent action token is based on the scene embedding tokens and previously generated tokens. Each action token is selected from a discrete action space of unique Verlet actions, which represent accelerations. A future trajectory is determined from the sequence of action tokens and provided to a motion planning module of the autonomous agent.

Abstract: Provided are methods for counterexample-guided update of a motion planner, which may include training a machine learning model to generate a trajectory for a vehicle. The training may generate a first trained machine learning model having a first plurality of feature weights. The machine learning model may be trained based on a counterexample for which the first trained machine learning model fails to generate a correct trajectory for the vehicle. The training may generate a second trained machine learning model having a second plurality of feature weights. A third plurality of feature weights may be determined by updating the first plurality of feature weights based on the second plurality of feature weights. The machine learning model may be updated to generate the trajectory of the vehicle by applying the third plurality of feature weights. Systems and computer program products are also provided.