Researchers led by Yansong Qu have presented SAFER, a framework that helps a self-driving car's world model foresee the outcome of rare, dangerous situations, according to a paper on arXiv. A world model is only reliable within the data it was trained on, and logs of good driving rarely contain the failures that would widen that coverage, the authors write.

Offline, partly trained agents and scripted disturbances create failures in simulation, and an expert labels every state visited with the right action and its value. When driving, a fast world model predicts consequences in real time, a planner scores candidate manoeuvres by their return minus a capped collision penalty, and the model keeps learning wherever the car leaves the offline data's coverage.

It led the compared methods on CARLA Leaderboard routes, with route completion of 0.95 and reward per step of 0.77 using a privileged bird's-eye view, and on MetaDrive, with episode rewards of 53.4 in normal and 49.0 in safety-critical traffic.

Source: arXiv Robotics (cs.RO)

Background

Yansong Qu

Robotics researcher

Robotics researcher.