---
language:
- en
tags:
- robotics
- dexterous-manipulation
- handson
- vision-language-action
- world-action-model
datasets:
- Cyan666/HandsOn-Datasets
papers:
- 2610.26765
---

# HandsOn: paper multi-task checkpoints

The eight multi-task checkpoints selected in **HandsOn: A Large-Scale Simulation Benchmark for Generalizable Dexterous Manipulation**.

**Upload complete.** All eight model payloads were verified against the source files. See `UPLOAD_VERIFIED.json` and `SHA256SUMS` for per-file verification.

[Paper](https://arxiv.org/abs/2610.26765) · [Code and inference adapters](https://github.com/TsinghuaC3I/HandsOn) · [Dataset](https://huggingface.co/datasets/Cyan666/HandsOn-Datasets) · [Model integration guide](https://github.com/TsinghuaC3I/HandsOn/blob/main/docs/paper-models.md)

| Model | Optimizer step | Checkpoint path |
|---|---:|---|
| π0.5 | 120,000 | `pi05/120000/` |
| DM0.5 | 86,000 | `dm05/checkpoint-86000/` |
| HY-Embodied-0.5-VLA | 125,000 | `hy/checkpoint-125000/` |
| LingBot-VLA | 151,000 | `lingbot/checkpoint-151000/` |
| OpenWAM | 70,000 | `openwam/step-070000/mp_rank_00_model_states.pt` |
| FastWAM | 100,000 | `fastwam/step-100000/mp_rank_00_model_states.pt` |
| Motus | 100,000 | `motus/step-100000/mp_rank_00_model_states.pt` |
| Spatial Forcing | 90,000 | `spatial/90000/` |

These are fixed paper selections, not checkpoints chosen by benchmark success. The single-task DP reference and the prompted GPT-6 controller are outside this weight release.

## Download

```bash
pip install -U huggingface_hub
# One model:
hf download Cyan666/HandsOn-Checkpoints --include 'pi05/**' CHECKPOINTS.json --local-dir ./handson-checkpoints
# All eight models:
hf download Cyan666/HandsOn-Checkpoints --local-dir ./handson-checkpoints
```

The native checkpoint formats and tensor bytes are preserved. Keep each Orbax `params/` tree and each DCP `state/` tree intact, including hidden metadata files. Standalone π0.5 optimizer state and OpenWAM RNG snapshots are omitted; native DCP and PyTorch containers can still contain training state. File sizes, model steps and relative paths are recorded in `CHECKPOINTS.json`.

## Inference

Use the matching `--model` profile from the [HandsOn model integration guide](https://github.com/TsinghuaC3I/HandsOn/blob/main/docs/paper-models.md). These are native research checkpoints, not `transformers.AutoModel` packages. The release's adapters preserve checkpoint-specific normalization, image preprocessing, denoising and action horizons.

π0.5 normalization is at `pi05/120000/assets/training_package_final18k.building/norm_stats.json`; Spatial Forcing normalization is at `spatial/90000/assets/v16_eef52/norm_stats.json`. The other adapters use the normalization assets packaged in the code release.

Foundation-model weights, tokenizers, and FastWAM/Motus language caches are supplied separately as documented in the integration guide. Obtain foundation assets from their original publishers and retain their terms. The code provides a language-cache preparation command.

Example, after installing the π0.5 inference environment from the code repository:

```bash
python scripts/serve_paper_policy.py --model pi05 \
  --checkpoint /path/to/handson-checkpoints/pi05/120000 \
  --normalization /path/to/handson-checkpoints/pi05/120000/assets/training_package_final18k.building/norm_stats.json \
  --tokenizer /path/to/paligemma_tokenizer.model
```

All profiles output absolute physical EEF52 commands, ordered right arm/hand then left arm/hand, for the shared 30 Hz native evaluator. Loading a checkpoint alone does not reproduce the paper's evaluation protocol or certify a new rollout.

## Provenance and terms

This repository publishes the selected HandsOn finetuning checkpoints in their original formats. Upstream model and component licenses continue to apply; this repository does not replace or broaden those terms. See the [native source notices](https://github.com/TsinghuaC3I/HandsOn/blob/main/docs/policy-source-notices.md) and the relevant foundation-model cards for component-specific information.

The upload verification records file sizes and SHA256 digests. It verifies storage and transfer integrity, not a fresh GPU evaluation or retraining.