README.md

---

pretty_name: RecoverBench
license: mit
task_categories:
  - robotics
tags:
  - robot-manipulation
  - error-recovery
  - mujoco
  - robosuite
---


# RecoverBench

> Camera-ready release v1.1.0: all 10,598 corpus trajectories have structurally readable action/state arrays. Two human files preserve their 17 complete CRC-verified NPY members but lack their original camera streams. The separate image supplement provides 880 explicitly aligned state-rendered frames; these are derivative observations, not recovered original pixels. See repair_manifest.json and IMAGE_SUPPLEMENT.md. Full-corpus action replay is not certified.

RecoverBench is a systematic benchmark for **error recovery in robotic manipulation**:
a taxonomy of 12 Error Skills × 2 difficulty degrees (24 subtypes, grouped into 5
Recovery Behavior Groups A–E), an automatic pipeline that injects errors into clean
trajectories and snapshots reproducible post-error states, and a recovery-demonstration
corpus for training recovery-capable policies. Tasks: `pick_place`, `stack`, `coffee`,
`threading`, `stack_three`, `three_piece_assembly` (robosuite / MuJoCo; mixed
Sawyer + Panda embodiments).

Code, evaluation protocol, and regeneration pipeline:
https://github.com/ErrorRecoveryBenchmark/ErrorRecoveryBenchmark

## Files

| File | Contents |
|---|---|
| `eval_data.tar` | **Evaluation set**: 1,360 error scenes (136 task × error-subtype combinations × 10 scenes). Each scene = `<id>.npz` (full MuJoCo state snapshot `post_sim_state`, quaternions in `(w,x,y,z)` order) + `<id>.json` (labels, environment fingerprint, RNG state). Layout: `error_scenes/<task>/scenes/`. |
| `training_scenes.tar` | **Training scene pool**: 5,038 error scenes generated by the same pipeline, disjoint from the evaluation set (verified: zero shared scene IDs and zero identical `post_sim_state` fingerprints). Recovery demonstrations were collected on scenes from this pool. Layout: `training_scenes/<task>/scenes/`. |
| `training_data.tar` | **Recovery demonstrations**: 973 readable human SpaceMouse teleoperation entries, including two without camera images (`recovery_demos_human/`) and MimicGen-style augmented demos (`recovery_demos_augmented/`, 8,957 trajectories). Each demo = `.npz` with `actions`, `states`, `eef_positions`, `target_poses`, `gripper_states`. |
| `recovery_mimicgen_generated.tar` | 668 generated recovery trajectories; part of the 10,598-entry corpus. |
| `recovery_training_success_from_error.tar` | Separate 739-trajectory HDF5 training artifact for the reported recovery-policy experiments. |
| `RecoverBench_image_supplement.zip` | State-rendered companion images for the two repaired records, original surviving image prefixes, provenance and validation. |
| `SHA256SUMS` | SHA-256 checksums for the five data archives and image supplement. |
| `eval_scene_ids.tsv` | The frozen evaluation-scene ID list (task ↔ scene id). |

## Quick start

```bash

# from the code repository

bash scripts/download_data.sh /path/to/data

export BENCHMARK_DATA=/path/to/data

```

Evaluation protocol (success = 10 consecutive success frames, 500-step budget,
single seed per scene) and metrics are specified in the code repository's
`EVALUATION.md`.

## Reproducibility notes

- Simulator stack is version-locked: robosuite 1.4.1 @ `c848ca84`, `mujoco==2.3.2`,
  `numpy<2.0` (see `ROBOSUITE_VERSION_LOCK.txt` in the code repo).
- Every error scene carries an environment fingerprint and RNG state for
  deterministic restoration via the code repo's replay tooling.
- Train/eval disjointness was verified at the level of scene IDs and SHA-256
  fingerprints of the canonical post-error simulator state.

## Citation

Zijia Tang, Ganlong Zhao, Xingping Chen, Junye Chen, Zihao Mo, Zedi Wang, Guanbin Li. RecoverBench: A Systematic Benchmark for Error Recovery in Robotic Manipulation. NeurIPS 2026, Evaluations and Datasets Track (accepted).

## License and provenance

The RecoverBench release is licensed under MIT; see LICENSE. Upstream simulator,
MimicGen source code, and seed datasets retain their own terms and attribution.
Seeds: https://huggingface.co/datasets/amandlek/mimicgen_datasets.

MimicGen: https://github.com/NVlabs/mimicgen (NVIDIA Source Code License).



Created and maintained by Zijia Tang, Ganlong Zhao, Xingping Chen, Junye Chen, Zihao Mo, Zedi Wang, Guanbin Li. Affiliations are in `CITATION.cff`. Correspondence: liguanbin@mail.sysu.edu.cn; zt81@duke.edu. Funding: NSFC 62322608 and Shenzhen Loop Area Institute FPF10120260001; compute: National Supercomputer Center in Guangzhou. The authors declare no competing interests; the submitting author confirmed all-coauthor agreement on October 21, 2026.

Use the code repository issue tracker for corrections.



## Machine-readable records



`eval_scenes.csv`, `training_scenes.csv`, and `recovery_demos.csv` index every NPZ
member of the core archives. `croissant.json` provides typed CSV column mappings,
archive checksums, RAI metadata, and upstream provenance. NPZ numerical arrays are
loaded separately with NumPy; the CSV indexes do not imply tensor decoding support.
The inventory verifies zero shared scene IDs and exact post-state fingerprints
between the frozen evaluation and training scene archives.

## Limitations and intended use

Simulation-only tabletop recovery evaluation and policy training; mixed Sawyer and
Panda embodiments, six tasks, hand-designed failure types, and uneven subtype
coverage. This is not evidence of safe real-world deployment or general-purpose
robot capability. Report clean-state and recovery success together.

The corpus contains 973 human, 668 generated, and 8,957 augmented trajectories (10,598 total). The separate 739 HDF5 trajectories are the recovery-policy training artifact, not additional corpus members.

## Image repair and reproducibility

Two truncated human NPZ containers were recovered without changing any complete numerical payload. Their original camera streams cannot be fully recovered. `RecoverBench_image_supplement.zip` contains companion image files with 381 and 499 state-aligned rendered frames, explicit state indices and timestamps, the surviving original image prefixes, provenance, and validation. These images use a consistent current pinned simulator camera; original viewpoint and crop timing could not be confirmed. Do not splice them into the original camera stream or describe them as original observations. No reported policy results were recomputed with these images. See [IMAGE_SUPPLEMENT.md](IMAGE_SUPPLEMENT.md).

The 739 HDF5 demos pass finite action/state and length checks. Their recorded model XML retains historical asset paths; normalize them with the pinned environment's `edit_model_xml` before replay. Structural checks do not certify action replay. Historical dataset commits remain available for comparison.

Five CSV indexes expose evaluation scenes, training scenes, human/augmented demos, generated demos and the separate training demos to Croissant. The image supplement is an additional FileObject, not extra trajectory records.