ActProbe: Action-Space Probe for Early Failure Detection of Generative Robot Policies

Abstract: Generative robot policies can fail unpredictably, while existing online detectors require internal model access or costly runtime resampling. ActProbe is a lightweight action-space detector that uses Temporal Consistency Error between consecutive action chunks and Action Chunk Magnitude from a single forward pass. A task-conditioned LSTM-MLP maps these signals to per-step failure probabilities. Across diverse policies and benchmarks, ActProbe improves the accuracy-timeliness Pareto frontier, transfers to unseen real-robot tasks, and accelerates PPO fine-tuning with 2.9x fewer environment interactions.

@article{actprobe,
  title={ActProbe: Action-Space Probe for Early Failure Detection of Generative Robot Policies},
  author={Bingjia Huang and Xiangyu Li and Xiang Wang and Liang Mi and Zixu Hao and Weijun Wang and Hao Wu and Kun Li and Yunxin Liu and Ting Cao},
  journal={arXiv preprint arXiv:2606.08508},
  eprint={2606.08508},
  archivePrefix={arXiv},
  url={https://arxiv.org/abs/2606.08508},

  year={2026}
}