Blame for README.md

---
pretty_name: LocoManiLab benchmark data, simulator assets, and checkpoints
license: mit
task_categories:
- robotics
tags:
- robotics
- humanoid
- loco-manipulation
- lerobot
- mujoco
- unitree-g1
viewer: false
---

# LocoManiLab release

Benchmark data, simulator assets, and fine-tuned policy checkpoints for
[LocoManiLab](https://github.com/Gamma-Robot/LocoManiLab), a workflow for humanoid
(Unitree G1, SONIC whole-body controller) loco-manipulation policy research. The code,
setup guides, and download tools live in the GitHub repository; this repository holds the
large files.

| Path | Contents | Size |
|---|---|---|
| `dataset/` | LeRobot v2.1 datasets, one directory per task | 33 GB |
| `asset/` | Asset manifest, task map, and the few asset files hosted here (`NOTICE.md`) | 18 MB |
| `checkpoint/gr00t/` | GR00T N1.7, fine-tuned on train50, step 10000 | 6.9 GB |
| `checkpoint/dm05/` | DM0.5, fine-tuned on train50, step 50000 | 23 GB |
| `checkpoint/xr1/` | XR-1 (Lightning run directory), fine-tuned on train50, step 50000 | 11 GB |
| `checkpoint/controller/` | G1 whole-body controller weights (SONIC, HEFT, AMO, ScaleBFM, ALMI) for `g1_backend` | 1.2 GB |

## Download

Use the helper from a LocoManiLab checkout (`pip install -e '.[data,hf]'`):

```bash
python scripts/hf/download.py list
python scripts/hf/download.py dataset --task IThorKitchenOpenFridge_scene13 --out datasets/hf
python scripts/hf/download.py assets --task IThorKitchenOpenFridge_scene13 --assets-dir external/humanoid_benchmark/robocasa/models/assets
python scripts/hf/download.py checkpoint gr00t --out checkpoints/hf
python scripts/hf/minimal_test.py --work hf-minimal-test
```

or the Hub CLI directly:

```bash
hf download GammaRobotics/LocoManiLab --repo-type dataset \
  --include "dataset/main/atomic/seen/IThorKitchenOpenFridge_scene13/*" --local-dir datasets/hf
```

See [data, assets, and checkpoints](https://github.com/Gamma-Robot/LocoManiLab/blob/main/docs/hf_resources.md)
for presets, checkpoint requirements, and the minimal download-and-test unit.

## Dataset

Every task directory is a LeRobot v2.1 dataset (`meta/`, `data/chunk-*/episode_*.parquet`,
`videos/chunk-*/observation.images.ego_view/episode_*.mp4`) recorded at 50 Hz in MuJoCo
with the G1 robot and the SONIC `v1_1` controller, plus three real-robot tasks.

| Group | Contents |
|---|---|
| `main/{atomic,composite}/{seen,unseen_scene,unseen_task}/` | the 234 benchmark tasks; names match `lab tasks` and `benchmark/tasks.yaml` |
| `data_generation/{human,gpt_aug,kimodo}/` | the same tasks from different motion sources |
| `multi_controller/{almi,amo,heft,scalebfm,sonic}/` | tasks recorded with other whole-body controllers |
| `real/`, `real2sim/` | three TeleLab tasks on the real robot and their simulated replicas |

Each frame has a 320×240 RGB ego view, a 43-D joint state (`observation.state`), end-effector,
root-orientation and projected-gravity observations, the task instruction, and actions
including the 64-D SONIC motion token and 7-D left/right hand targets. Directories under
`data_generation/kimodo/` and two under `real/` have no `meta/stats.json`; LocoManiLab's
`lab validate` and `lab prepare` do not need it.

## Simulator assets

Apart from the files listed in `asset/NOTICE.md`, this repository does not redistribute simulator assets. `asset/manifest.tsv.gz`
lists every file the benchmark tasks load (path relative to
`humanoid_benchmark/robocasa/models/assets`, sha256, size, source) and `asset/tasks.json.gz`
maps each task to its files. The download helper fetches each file from its original host
and verifies its checksum; all 234 benchmark tasks plus the randomization texture pool need
22,745 files (3.4 GB), one iTHOR task about 32 MB.

| Source | Content | Terms |
|---|---|---|
| [MolmoSpaces](https://huggingface.co/datasets/allenai/molmospaces) | iTHOR, ProcTHOR, Holodeck scenes; THOR and Objaverse objects | CC BY 4.0; Objaverse subset ODC-BY 1.0, per-object licenses |
| [RoboCasa](https://github.com/robocasa/robocasa) asset archives | textures, generative textures, Objaverse and AI-generated object categories | CC BY 4.0 |
| [NVIDIA PhysicalAI Kitchen MJCF](https://huggingface.co/datasets/nvidia/PhysicalAI-Robotics-Manipulation-Objects-Kitchen-MJCF) | Lightwheel fixtures and objects | CC BY 4.0 |
| simulator checkout | TeleLab scenes and objects, layouts, task specs | the simulator's terms |
| `asset/local.tar.gz` (this repository) | Gamma Robotics' converted box (`objects/usd/box/converted`) | MIT |
| `asset/local.tar.gz` (this repository) | a folder from [ArtVIP](https://huggingface.co/datasets/X-Humanoid/ArtVIP) `folder_25`, materials and joints modified, converted to MuJoCo (`fixtures/articulated/green_folder001`) | Apache-2.0 (`asset/LICENSE-Apache-2.0.txt`) |
| `asset/local.tar.gz` (this repository) | a RoboCasa object file with a corrected texture path | CC BY 4.0 (RoboCasa) |

## Checkpoints

Inference weights only; optimizer states were not released, so the checkpoints cannot
resume training. Serving instructions are in each policy guide of the GitHub repository.

- `gr00t/`: serve with the Isaac-GR00T environment. It loads the gated backbone
  `nvidia/Cosmos-Reason2-2B` at startup (accept its terms and log in on the serving machine).
- `dm05/`: serve with the OpenDM environment.
- `xr1/`: serve with the Xiaomi-Robotics-1 environment and a local Qwen3-VL cache; the run
  directory has no `last.ckpt`, so pass `--ckpt-step 50000`.

### Whole-body controller weights

`checkpoint/controller/` mirrors `<g1_backend>/checkpoints/` and holds only the ONNX and
TorchScript weights; the matching `policy.yaml` manifests are tracked in
[g1_backend](https://github.com/LocoManiLab/g1_backend). Fetch only the controller you run
with the simulator's helper, for example the SONIC `v1_1` controller used by the benchmark data:

```bash
python external/humanoid_benchmark/robocasa/scripts/download_g1_checkpoints.py \
  --family sonic_v1_1 --dest external/g1_backend/checkpoints
```

| Family | Path | Origin |
|---|---|---|
| `sonic` (low-latency, with the shared planner) | `sonic/low_latency/g1/` | [NVIDIA GEAR-SONIC](https://huggingface.co/nvidia/GEAR-SONIC) |
| `sonic_v1_1` | `sonic/v1_1/g1/` | [NVIDIA GEAR-SONIC](https://huggingface.co/nvidia/GEAR-SONIC) |
| `heft` | `heft/pmg/` | HEFT, converted for g1_backend |
| `amo` | `amo/` | AMO, converted for g1_backend |
| `scalebfm` | `scalebfm/humanoid_transformer_m/` | ScaleBFM, converted for g1_backend |
| `almi` | `almi/flat_dual/` | mjlab `Mjlab-Velocity-Flat-Unitree-G1-Dual`, exported to ONNX |

## Sources and terms

The data recorded by this project, the fine-tuned checkpoints, and Gamma Robotics' own
assets are released under the MIT License (`LICENSE`). Third-party material keeps its own terms, which
apply in addition:

- Simulator assets are fetched from their original hosts under those hosts' terms (see
  above); the benchmark data renders them. Two of the 47 Objaverse objects the tasks load
  are licensed CC BY-NC 4.0: framed wall art `596fd518055a49a28d56c193cd087470` (by 2zqa)
  and `f33af1b399a24448880930fef312ef3f` (by jamesgourge) in Holodeck scene 117. They are
  fetched from MolmoSpaces, not hosted here, and may be visible in the scene-117 videos;
  their non-commercial terms apply to that content.
- Controller weights in `checkpoint/controller/` are converted deploy artifacts of the
  upstream projects listed above and keep those projects' terms; the SONIC weights are under
  the NVIDIA Open Model License.
- Checkpoints are fine-tuned from `nvidia/GR00T-N1.7-3B` (with `nvidia/Cosmos-Reason2-2B`),
  DM0.5, and `XiaomiRobotics/Xiaomi-Robotics-1-5B` (with `Qwen/Qwen3-VL-4B-Instruct`); their
  model licenses and acceptance requirements apply.

Check the terms of each source before redistributing or using this material commercially.