Behavior Prediction from Sensor + Camera Data

Requirements Inputs: synchronized LiDAR/radar returns and camera streams from the ego vehicle. Outputs: a trajectory forecast for every surrounding agent over a future horizon of roughly 3–8 seconds, along with uncertainty expressed either as per-step covariance or as a categorical distribution over discrete modes. Downstream contract: the predictions are consumed by the motion planner. The planner needs (a) deterministic output at a fixed cadence, (b) uncertainty in a form…