AVM-Eval

AVM-Eval (Action and Video Mismatch Benchmark) evaluates how reliably inverse dynamics models and reward models assess action-following fidelity on real and generated robot videos. It is introduced in ACWM-RM: A Reference-Free Reward Model from Complementary Mismatch Preferences.

Dataset Overview

AVM-Eval contains approximately 67.2k evaluation pairs:

Pair type	Size	Description
Action mismatch	53.9k	The same recorded video paired with original and perturbed action sequences, covering numerical and semantic perturbations.
Video mismatch	13.3k	Two generated videos conditioned on the same initial frame and action sequence, covering intra-model and cross-model comparisons. Preferences are determined using nDTW-based action-following scores.

Source data comes from AgiBotWorld, RoboTwin, RoboDojo, and Genie-Sim. Generated videos are produced by DreamDojo-2B, DreamDojo-14B, GE-Sim-2B, and OSCAR-2B. The benchmark uses pairwise accuracy to evaluate action-following assessment, including generalization to unseen tasks, embodiments, and scenes.

Files

action_mismatch/   # Action-mismatch evaluation pairs in Parquet format
video_mismatch/    # Video-mismatch evaluation pairs in Parquet format
media.tar          # Media archive

See the paper for construction details and the evaluation protocol.

Citation

If you use AVM-Eval in your research, please cite:

@misc{hong2026acwmrm,
  title         = {{ACWM-RM}: A Reference-Free Reward Model from Complementary Mismatch Preferences},
  author        = {Shibo Hong and Siyu Ren and FengJiao Chen and Zhongyuan Peng and Wei Shi and Yixin Cao},
  year          = {2026},
  eprint        = {2610.32222},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2610.32222}
}