FailTime

Failure annotations for robot manipulation rollouts, from the paper ProTracer: Proprioception-Guided Failure Diagnosis in Robot Manipulation. Each episode is labelled as a success or a failure, and every failure carries the moment it went wrong, its failure type, the reason, and how it could have been avoided.

• FailTime-Short — annotations for ViFailback recordings (AgileX Piper). Download ViFailback separately for the episodes themselves.
• FailTime-Long — long-horizon bimanual recordings (ALOHA and SO-101), included here as self-contained HDF5 files.

Layout

failtime_short/    train.json  test.json  tasks/
failtime_long/     test.json   tasks/     episodes/*.hdf5

tasks/ holds one XML per task: instruction, objects and substages.

Annotations

{"sample": "pour_ball_transfer_cup_episode_19", "task_id": "pour_ball_transfer_cup", "episode": "episode_19",
 "robot": "aloha", "hdf5": "episodes/pour_ball_transfer_cup_episode_19.hdf5", "verdict": "fail",
 "onset": 1.146, "type": ["wrong_object"], "reason": "Expected: ...", "avoidance": "..."}

onset (seconds), type, reason and avoidance are given for failures only.

Episodes

Every FailTime-Long HDF5 holds one episode:

observation.state_eef	both end-effectors: xyz, quaternion xyzw, gripper
observation.state, action	joint states and actions (observation.effort on ALOHA)
observation.images.<camera>/mp4_bytes	three camera videos: left wrist, right wrist, overhead
timestamp, frame_index, episode_index, next.done	frame bookkeeping

Usage

hf download ChangUoA/FailTime --repo-type dataset --local-dir data
ln -s /path/to/ViFailback/raw_data data/failtime_short/episodes

The code is at github.com/Chang-AIML/ProTracer.

Citation

@article{dong2026protracer,
  title  = {ProTracer: Proprioception-Guided Failure Diagnosis in Robot Manipulation},
  author = {Dong, Chang and Hosseinzadeh, Mehdi and Wong, King Hang and Liu, Lingqiao and Fraysse, Francois and Dayoub, Feras and Nguyen, Minh Hoai},
  year   = {2026}
}