Quentin Rolland^1,2, Fabrice Mayran de Chamisso¹, Jean-Baptiste Mouret^2,3

¹Université Paris-Saclay, CEA, List, F-91120, Palaiseau, France, ²Inria, CNRS, Université de Lorraine, LORIA, F-54000 Nancy, France, ³Bleu Robotics, Paris, France

IEEE International Conference on Robotics and Automation (ICRA), 2026

The official code repository for "Failure Identification in Imitation Learning via Statistical and Semantic Filtering," presented at ICRA 2026.

Imitation Learning (IL) policies are brittle to rare or out-of-distribution events in real-world robotic deployments.
We introduce FIDeL, a policy-agnostic failure identification framework that combines:

• Vision-based anomaly detection
• Optimal transport alignment with expert demonstrations
• Spatio-temporal thresholding via conformal prediction
• Semantic filtering using Vision-Language Models (VLMs)

FIDeL detects, localizes, and semantically filters failures in real time, without interfering with policy execution.

We also introduce BotFails, a multimodal dataset for robotic failure detection:

• Vision, proprioception, and language instructions
• 646 video sequences
• 414,359 annotated frames
• Real-world manipulation and interaction tasks
• Explicit failure and benign anomaly annotations

FIDeL outperforms state-of-the-art anomaly detection baselines on BotFails, achieving:

• +5.30% AUROC in anomaly detection
• +17.38% accuracy in failure identification

Qualitative results and videos are available on the project webpage.

If you find FIDeL useful, please consider citing our work:

@inproceedings{rolland2026failure,
  title={Failure Identification in Imitation Learning via Statistical and Semantic Filtering},
  author={Rolland, Quentin and Mayran de Chamisso, Fabrice and Mouret, Jean-Baptiste},
  booktitle={IEEE International Conference on Robotics and Automation (ICRA)},
  year={2026},
}

Parts of this project page were adopted from the Nerfies page.

This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.