---
license: apache-2.0
size_categories:
- 1K<n<10K
task_categories:
- robotics
pretty_name: 'hero_bench_v1: HERO humanoid reaching benchmark'
tags:
- humanoid
- whole-body-control
- end-effector-tracking
- unitree-g1
- mujoco
- benchmark
---

# hero_bench_v1 — HERO humanoid end-effector reaching benchmark

[Paper](https://huggingface.co/papers/2602.16705) | [Project](https://hero-humanoid.github.io/) | [Code](https://github.com/RunpeiDong/HERO)

`hero_bench_v1` is the public end-effector reaching benchmark of **HERO** (Learning Humanoid End-Effector Control for Visual
Whole-Body Open-Vocabulary Object Grasping, CoRL 2026): 1,298 synthetic whole-body reaching reference motions for the
Unitree G1 humanoid with the Dex3 hand, scored in MuJoCo with the tooling of the HERO release
(`scripts/hero_bench.py`, `sim2sim/bench/`). The paper's 180-clip protocol (three table heights × 60 targets, 3 s hold) is contained
byte for byte as the core tier's `paper` layer (`h050` / `h074` / `h088`).

| tier | clips | what it covers |
|---|---|---|
| core | 420 | paper protocol (180) + reaches to 0.60 m, cross-body targets, 0.62 / 0.80 m tables, top-down and tilted grasps |
| extended | 618 | 0.30 / 0.40 m tables, floor pick-up, 1.00 / 1.10 m shelves, wide lateral targets, palm-down / palm-up / fanned grasps, hovering targets, half-speed reaches, 6 s holds, retraction, table-edge targets |
| stress | 260 | bowing reaches to 0.70 m, a 0.25 m table, 1.15 / 1.20 m shelves, 1.33× speed, and a pool of clips with 1.5–3 cm IK terminal residual |

Each clip is a 50 Hz reference (`.npz`): 0.3 s settle at the rest pose, a 1.5–6 s reach to a world-fixed palm target, a 3 s hold
(6 s in `hold6`), and for the `retract` layer a return to the rest pose. `BENCH_MANIFEST.json` lists every clip with its target
pose, frame indices (`reach_end_frame`, `hold_end_frame`, `retract_start_frame`), acceptance level and IK residual, plus the frozen
evaluation protocol (`protocol` block, hashed in `protocol_sha256`). `BENCH_REPORT.md` is the per-layer summary.

## Files

```
hero_bench_v1_corpus_33542601.tar.gz            # the frozen corpus (about 520 MB; 33542601 = first 8 hex of its sha256, full hash in the .sha256 file next to it)
hero_bench_v1/                                # the same corpus unpacked, for browsing / partial download
  BENCH_MANIFEST.json  BENCH_REPORT.md  SHA256SUMS  DATA_LICENSE  tiers/
  core/  extended/  stress/                   # <stratum>__<clip_id>.npz + per-tier BENCH_MANIFEST.json
```

## Usage

```sh
git clone https://github.com/RunpeiDong/HERO && cd HERO && pip install -e ".[bench]"
python scripts/hero_bench.py fetch --url https://huggingface.co/datasets/RunpeiDong/hero_bench/resolve/main/hero_bench_v1_corpus_33542601.tar.gz \
    --sha256 33542601ed456ac606f8177ab75892d2e616585ec562f01976cb6af238c9c5d0 --out data/hero_bench_v1
python scripts/hero_bench.py verify --dir data/hero_bench_v1
python scripts/hero_bench.py run --corpus data/hero_bench_v1 --onnx-dir checkpoints/example --sidecar checkpoints/example/model_hero.json \
    --out results/example --tiers core            # all = core,extended,stress
python scripts/hero_bench.py report --corpus data/hero_bench_v1 --results results/example --card
```

A result is comparable to published numbers when `verify` reports every file byte-identical and the same `protocol_sha256`.
Scores are world-frame palm errors of the reaching hand, open loop and with HERO's closed-loop replanning, with success rates
S7.5 / S5 (open loop) and C3 (closed loop) and Wilson 95 % intervals; see `docs/benchmark.md` in the code release for the protocol.

## Provenance

The clips are generated by the release's whole-body IK reach generator (`data_tools/hero_reach_generator.py`) from the plan in
`data_tools/reach_specs.py` (seeds, quotas, acceptance levels) and re-timed deterministically (`data_tools/retime_bench.py`);
no human motion-capture data is included. The corpus can be rebuilt with `scripts/hero_bench.py build`; rebuilds are
protocol-equivalent (same targets, frames and acceptance levels) but not guaranteed byte-identical across MuJoCo / solver builds.

## License

Apache License 2.0 (see `DATA_LICENSE`); the code release is MIT-licensed separately.

## Citation

```bibtex
@inproceedings{dong2026hero,
  title     = {{HERO}: Learning Humanoid End-Effector Control for Visual Whole-Body Open-Vocabulary Object Grasping},
  author    = {Dong, Runpei and Li, Ziyan and Gupta, Arjun and He, Xialin and Gupta, Saurabh},
  booktitle = {10th Annual Conference on Robot Learning},
  year      = {2026},
  url       = {https://openreview.net/forum?id=gbchkYm28k}
}
```