In plain words
A driving log usually records one path through a scene, even though several paths may be possible. That makes it hard to tell whether a model follows an instruction or simply copies the recorded path.
Cartender is a driving model that creates several candidate paths, checks whether the scene supports the instruction, and ranks the candidates. On fixed recorded inputs, it fulfilled both requests in a same-scene pair more often than a model trained only on matched recordings. The result does not show how it performs in interactive traffic.
The research
When a driving log shows only one path through a scene, how can we tell whether a model follows an instruction or merely matches the recorded path? The paper asks whether training on alternative paths for the same scene can help test and improve that behavior.
Cartender creates alternative paths for the same driving scene and trains a policy to use scene-supported instructions when choosing among them.
Cartender creates alternative paths for the same recorded scene, pairs them with instructions, and trains a driving policy to generate and rank candidate paths. The researchers then change instructions on fixed recorded inputs and judge behavior against action criteria.
Across three seeds, Cartender fulfills both requests in 57.00% of maneuver pairs, versus 23.67% for matched-only imitation with the same backbone.
On the expanded same-scene training set, Cartender also exceeds same-scene imitation on maneuver and qualifier paired success. This supports a contribution beyond those added examples under the reported procedure. Each pair requires both requests to be fulfilled, and missing outputs count as failures.
These results come from evaluations on fixed recorded inputs. Rule-based checks and RFS do not certify collision freedom or instruction fulfillment, and the compatibility gate still allows some unsupported requests to be followed. The study does not establish performance in interactive traffic.
Analysis
Deep
- Original paper authors
- Zhiwen Chen, Benjin Zhu, Yan Xie, Jun Ma
- Submitted to arXiv
- License
- CC BY 4.0