Event
Cartender Improves Instruction Following with Unselected Driving Branches
1 reports · 1 sources
Zhiwen Chen et al. propose the Cartender method, which uses Counterfactual Branching to construct matched and mismatched instruction-action pairs from real driving logs without a simulator: a vision-language model proposes alternative futures and their selection instructions, and instructions not supported by the scene serve as negative samples. On WOD-E2E, the best of four candidates achieves an RFS of 8.642 under 3,603 feasible instruction conditions; compared with matched-only imitation using the same backbone network, Cartender raises the success rate on same-scene paired instructions from 23.67% to 57.00%, and RFS from 6.32 to 7.10.
Report timeline
- Cartender: Verifiable Language-Conditioned Driving by Learning the Drives Not Taken
Cartender proposes Counterfactual Branching, which constructs matched and mismatched instruction-action pairs from real driving recordings without a simulator, using a vision-language model to propose alternative futures and their selecting instructions, with instructions unsupported by the scene serving as negative samples.
arXiv cs.RO ·