CoLT-Drive

CoLT-Drive: Counterfactual Long-Tail Benchmarking and Knowledge-Preserving Adaptation for Driving Affordance Prediction

Findings of the Association for Computational Linguistics: EMNLP 2026

Zhengxu Tang, Guofeng Cui, Ziyu Gong, Xiaozhou Zhang, Ruifeng Deng, Chengzhi Qi, Ke Chen, Sachin Patil, Tianjun Xiao, Langechuan Liu, Pichao Wang

🚧 Dataset Release Coming Soon

This repository will host the complete CoLT-Drive benchmark, including:

• All 3,536 counterfactual benchmark images (1,768 paired $v_{full}$ / $v_{clean}$ versions)
• Acceptable action-pair annotations (three-reviewer consensus labels)
• Data splits and per-sample metadata (base scene, obstacle type, position, affordance category, ego-motion history, navigation command)
• The annotation guidelines used by the labeling team

We are finalizing the release. Please check back soon. Training and evaluation code will be released on GitHub: 👉 https://github.com/TangZhengxu/CoLT-Drive

About CoLT-Drive

CoLT-Drive is a counterfactual long-tail driving benchmark for decision-level affordance prediction: rare objects are inserted into otherwise fixed driving scenes, and models must predict how the inserted object changes the ego vehicle's feasible high-level longitudinal–lateral actions. The benchmark contains 3,536 reviewed samples built from 29 base scenes (9 from Alpamayo-R1, 20 from nuPlan mini) and 50 obstacle types across five affordance categories: living entities, nonliving entities, road hazards, full blockages, and false positives.

Labels denote the immediate high-level response under the current observation (e.g., cautious deceleration on approach to a distant blockage), not the vehicle's terminal maneuver. No relabeling was performed for the camera-ready; the paper's wording was only clarified to match this annotation semantics.

Citation

@inproceedings{tang2026coltdrive,
  title     = {CoLT-Drive: Counterfactual Long-Tail Benchmarking and Knowledge-Preserving Adaptation for Driving Affordance Prediction},
  author    = {Tang, Zhengxu and Cui, Guofeng and Gong, Ziyu and Zhang, Xiaozhou and Deng, Ruifeng and Qi, Chengzhi and Chen, Ke and Patil, Sachin and Xiao, Tianjun and Liu, Langechuan and Wang, Pichao},
  booktitle = {Findings of the Association for Computational Linguistics: EMNLP 2026},
  year      = {2026}
}

License

To be announced with the release.