File size: 5,281 Bytes
4c6e1bb 67f5a76 4c6e1bb 67f5a76 4c6e1bb 67f5a76 4c6e1bb 67f5a76 4c6e1bb 67f5a76 4c6e1bb | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 | ---
pretty_name: BiGym 2.0 Coding-Agent Rollouts
license: cc-by-4.0
language:
- en
task_categories:
- robotics
tags:
- robotics
- bigym
- humanoid
- loco-manipulation
- coding-agents
- evaluation
---
<p align="center">
<img src="https://huggingface.co/datasets/SWIRL-Lab/bigym2-agent-rollouts/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/bigym2-agent-rollouts/resolve/main/assets/bigym-lockup-dark.svg" alt="BiGym 2.0" width="320" class="hidden dark:block">
</p>
<h1 align="center">BiGym 2.0 — Coding-Agent Rollouts</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>
Every evaluation episode behind the coding-agent results of **BiGym 2.0**:
the frozen program from each development session, replayed on all 100 hidden
evaluation seeds. Two agents (Claude Opus 5.5 via Claude Code, GPT-6 Astra via
Codex CLI), nine tasks, three sessions each: 54 programs, 5,400 episodes. Each
replay reproduces the scored outcome of its seed.
## Layout
```
manifest.csv one row per episode
episodes/<task>/metadata.json the environment the task's episodes ran in
episodes/<task>/<session>_seed<seed>.npz simulator state and actions
viser/<task>/<session>_seed<seed>.viser 3D replay for the viser player
viser/<task>/<session>_seed<seed>.viser.json its frame count, length and camera
trace/<task>/<session>_seed<seed>.json lines the program ran at each control step
```
Sessions are named `opus-s1` … `opus-s3` and `astra-s1` … `astra-s3`; seeds run
from 620000 to 620099.
## Watching them
Each `episodes/<task>/` folder opens in the BiGym 2.0 viewer, with every
episode's seed, outcome and length in its list:
```bash
hf download SWIRL-Lab/bigym2-agent-rollouts --repo-type dataset --include "episodes/move_plate/*" --local-dir bigym2-agent-rollouts
uv run bigym-view --demo-dir bigym2-agent-rollouts/episodes/move_plate
```
`--include "episodes/move_plate/astra-s1_*"` fetches one session only. The
viewer is in the [code repository](https://github.com/swirl-uk/BiGym2).
## Episodes
One `.npz` per episode, stepped at the 50 Hz control rate (step *i* is at
*i*/50 s). No images are stored: the full simulator state is, so any camera
at any resolution can be re-rendered.
| Key | Shape | |
|---|---|---|
| `full_qpos` | (T+1, nq) | positions of the robot and every object |
| `full_qvel` | (T+1, nv) | velocities |
| `action` | (T, 20 or 21) | joint targets the program sent |
| `reward` | (T, 1) | per-step reward |
| `seed`, `length` | scalar | evaluation seed, T |
| `success`, `fell` | scalar | 1.0 if the goal state held for one second / if the robot fell |
| `termination` | scalar | `success`, `timeout`, `fell`, `physics_error` or `terminated` |
`nq` and `nv` depend on the task's objects. `reach_target_*`, `move_plate` and
`drawer_top_*` use a 20-dim action; the other tasks add a torso-pitch command.
`metadata.json` holds the environment configuration (`env_config`), which
`EnvConfig.from_metadata` reads back.
## Code traces
Which lines of its program each episode ran, step by step: the episode was
re-run with the frozen program under Python's `sys.settrace`, and a trace is
kept only if that re-run's state trajectory is identical to the scored one,
frame for frame. 5,347 of the 5,400 episodes have one. The other 53 (43 of
them `reach_target_multi_modal/opus-s1`) re-run with float drift from about
step 42: same length and outcome, up to a few millimetres apart.
Line numbers refer to the program as the project page lists it: `policy.py`,
then each helper module in name order, each after three extra lines (a blank line, a
`# ──────── <file> ────────` divider, a blank line).
| Key | |
|---|---|
| `length`, `success` | episode length T, outcome |
| `sets` | the distinct sets of lines run in one step |
| `steps` | for each of the T steps, the index of its set in `sets` |
| `window`, `segments` | for display: phases of the episode, as `[from, to, line, [methods]]` over runs of the lines run in the last `window` steps |
## Manifest
`manifest.csv` has `task`, `session`, `model`, `harness`, `program_version`
(the development iteration that was frozen and scored), `seed`, `success`,
`length`, and the paths of the episode and its replay.
## Related
The human VR demonstrations of the same tasks are in
[`SWIRL-Lab/bigym-g1-native60`](https://huggingface.co/datasets/SWIRL-Lab/bigym-g1-native60).
## 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}
}
```
|