---
license: other
license_name: mixed
tags:
  - robotics
  - vision-language-action
  - hand-pose
---

# VLA-HAND baselines

Checkpoints of the two external baselines trained against V4 in the VLA-HAND project.
Both were trained and tested on the same three domains (GigaHands, OakInk2, TACO), splits,
normalisation statistics and test protocol as V4. Each is the checkpoint with the lowest
validation loss, the one all reported test numbers come from.

| folder | model | init | action space | step | license |
|---|---|---|---|---|---|
| `vitra_original/` | PaliGemma2-3B + DiT/DDPM action head (VITRA) | `VITRA-VLA/VITRA-VLA-3B` | MANO joint angles | 10,000 | Gemma Terms of Use, see `vitra_original/NOTICE.md` |
| `smolvla_official/` | SmolVLA (LeRobot `SmolVLAPolicy`) | SmolVLM2-500M | 21-keypoint hand pose | 18,000 | Apache-2.0 |

## Test results

Test split, the first 8 control steps of each test window, three domains pooled.
Wrist translation in cm, wrist rotation in degrees (geodesic), fingers in cm; lower is better.

| model | wrist translation | wrist rotation | fingers |
|---|---|---|---|
| SmolVLA | 0.7518 | 3.6235 | 1.0094 |
| VITRA | 1.0298 | 4.7789 | 1.7933 |

VITRA predicts joint angles, so its finger error is measured after MANO forward kinematics
and includes an FK floor that SmolVLA does not have. The finger columns are not directly
comparable.

## Files

Each folder holds `best.pt` and `training_config.json`, the run config stored inside the
checkpoint (its paths refer to the training machine). `best.pt` is a `torch.save` dict with:
`step`, `model` (state dict), `config`, `val` (validation metrics at that step), and
`action_mean` / `action_std`, the normalisation statistics the model was trained with.
There is no optimiser state.

## Code

Training and test code: https://github.com/kaichen-z/VLA-HAND/tree/Kai_final/yuzhench_distill_editor/baselines
(commit `baee56f`). Load with the matching package's `Model` and `load_pretrained(path)`.

## Full T-ReX human-action checkpoint

[model/trex_full_egotouch](model/trex_full_egotouch) contains all inference weights, the processor, and standalone loading code.
No separate upstream checkpoint download is required. This is the no-touch EgoTouch adaptation, not native tactile T-ReX.
Its checkpoint uses `human_trex.full.v1`, not the legacy baseline format described above.

## Tachin adaptation checkpoints

[model/tachin_adaptation_20260922](model/tachin_adaptation_20260922) contains validation-selected SmolVLA, VITRA, and HumanTReX checkpoints after Tachin adaptation.
These are trained-parameter checkpoints, not standalone full models. The release documents their frozen initialization dependencies.
No tactile editor or full test results are included. The existing three-domain and EgoTouch releases remain unchanged.

## EgoTouch T-ReX editor replica

[model/trex_editor_egotouch_replica_20260930](model/trex_editor_egotouch_replica_20260930) contains the validation-selected tactile editor from the September 30 video-generation replica. This is not the original paper-run checkpoint. It is separate from the full HumanTReX backbone. The package includes a strict loader and a verification record.