Reproduce “Cartender: Verifiable Language-Conditioned Driving by Learning the Drives Not Taken” 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
Zhiwen Chen, Benjin Zhu, Yan Xie, Jun Ma
Cartender tests instruction control by creating alternative instruction–trajectory examples from real recordings, without simulator rollouts, and evaluating requests on identical recorded inputs. Its compatibility score moderates language conditioning, while reward refinement trains instruction-sensitive ranking. On WOD-E2E, maneuver paired success—both feasible requests passing action and hard-constraint checks—was 57.00%, versus 23.67% for matched-only imitation; evaluation was not closed-loop.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.