Reproduce “ChauffeurNet: Learning to Drive by Imitating the Best and Synthesizing the Worst” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Mayank Bansal, Alex Krizhevsky, Abhijit Ogale

ChauffeurNet trains a driving model to recover from departures from logged routes. It perturbs trajectories in a bird’s-eye scene and combines imitation learning with collision and off-road penalties. In closed-loop simulation, final model M4 passed in 90% of 20 parked-car scenarios, versus 55% for M3; this result is limited to the constructed passing task.

Key insights

  • Pure behavior cloning, even with extensive expert data and mid-level representations, is insufficient for robust closed-loop autonomous driving due to distribution shift.
  • Generating synthetic "worst-case" perturbations and combining them with specific environment losses can effectively teach avoidance behaviors and improve robustness without active online interaction.
  • Closed-loop simulation and real-world testing are essential for evaluating driving policies, as open-loop metrics can be misleading and do not reflect real-world performance.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.