---
pretty_name: BiGym 2.0 G1 Demonstrations
license: cc-by-4.0
language:
- en
task_categories:
- robotics
tags:
- robotics
- lerobot
- bigym
- humanoid
- loco-manipulation
- imitation-learning
- teleoperation
---

<p align="center">
  <img src="https://huggingface.co/datasets/SWIRL-Lab/bigym-g1-native60/resolve/main/assets/bigym-lockup-light.svg" alt="BiGym 2.0" width="320" class="block dark:hidden">
  <img src="https://huggingface.co/datasets/SWIRL-Lab/bigym-g1-native60/resolve/main/assets/bigym-lockup-dark.svg" alt="BiGym 2.0" width="320" class="hidden dark:block">
</p>

<h1 align="center">BiGym 2.0 — G1 Demonstrations</h1>

<p align="center">
  <a href="https://arxiv.org/abs/2610.07594"><img alt="arXiv" src="https://img.shields.io/badge/arXiv-2610.07594-b31b1b.svg"></a>
  <a href="https://github.com/swirl-uk/BiGym2"><img alt="Code" src="https://img.shields.io/badge/GitHub-BiGym2-181717?logo=github"></a>
  <a href="https://creativecommons.org/licenses/by/4.0/"><img alt="License" src="https://img.shields.io/badge/License-CC_BY_4.0-lightgrey.svg"></a>
</p>

Human VR demonstrations for the 20 tasks of **BiGym 2.0**: 60 successful
episodes per task on a Unitree G1 (Dex1 grippers), recorded through the same
frozen GR00T-WBC whole-body controller used for evaluation.

![BiGym 2.0 tasks](assets/teaser.jpg)

## Format

- One **LeRobot v3** dataset per task folder, 50 Hz.
- Three cameras (`head`, `left_wrist`, `right_wrist`) at 84×84, lossless PNG.
- Per-frame simulator state (`full_qpos`, `full_qvel`), so any camera or
  resolution can be re-rendered (`bigym-rerender-lerobot`).
- English task instruction in each folder's `meta/tasks.parquet`.

## Tasks

| Group | Tasks |
|---|---|
| Reaching | `reach_target_single`, `reach_target_multi_modal`, `reach_target_dual` |
| Table-top | `move_plate`, `move_two_plates`, `flip_cup`, `flip_cutlery`, `stack_blocks` |
| Dishwasher | `dishwasher_close`, `dishwasher_load_cups`, `dishwasher_load_cutlery`, `dishwasher_load_plates` |
| Kitchen counter | `drawer_top_open`, `drawer_top_close`, `pick_box`, `put_cups`, `saucepan_to_hob`, `sandwich_remove`, `wall_cupboard_open`, `wall_cupboard_close` |

**Action width differs by task.** `reach_target_*`, `move_plate` and
`drawer_top_*` use a 20-dim action / 50-dim state; the other 14 tasks add a
torso-pitch command and use 21-dim action / 56-dim state. Multi-task training
must reconcile the two.

## Usage

```bash
hf download SWIRL-Lab/bigym-g1-native60 --repo-type dataset --include "move_plate/*"
```

Each task folder is a standalone LeRobot v3 dataset. Environments and loaders are in the
[code repository](https://github.com/swirl-uk/BiGym2).

## Related

Every evaluation episode of the paper's coding-agent sessions, 5,400 in all,
with simulator state, actions and the program lines each step ran, is in
[`SWIRL-Lab/bigym2-agent-rollouts`](https://huggingface.co/datasets/SWIRL-Lab/bigym2-agent-rollouts).

## Citation

```bibtex
@article{zhang2026bigym2,
  title   = {BiGym 2.0: Benchmarking Learned and Agent-Developed Policies for Humanoid Household Manipulation},
  author  = {Zhang, Zexi and Zhu, Zecheng and Chen, Zidong and Tuya, Zulkhuu and James, Stephen},
  journal = {arXiv preprint arXiv:2610.07594},
  year    = {2026}
}
```