README.md

---
license: other
tags:
  - robotics
  - vision-language-action
  - hand-pose
  - egotouch
---

# VLA-HAND Table 1 checkpoints

Adapted checkpoints for the EgoTouch Table 1 protocol. The model folder follows the layout used by `MIT-Media-Lab/vla-hand-baselines`.

| folder | model | dataset | action space |
|---|---|---|---|
| `smolvla_egotouch/` | SmolVLA | EgoTouch | 21-keypoint hand pose |

## Paired 10 Hz / 30 Hz editor cache (2026-09-21)

The [table1_editor_cache_paired30_v2_20260921](https://huggingface.co/MIT-Media-Lab/vla-hand-table1/tree/main/table1_editor_cache_paired30_v2_20260921) folder contains the revised, common cache for SmolVLA, VITRA, and ten-domain T-WLA.

All three models use the same 415,624 train, 23,571 validation, and 23,950 test plans.
Each plan supports causal tactile updates at 0, 33.333, and 66.667 ms.
Action labels use source-frame time. T-WLA memory retains its 10 Hz update cadence.

The release includes arrays, train-only statistics, source provenance, audits, checksums, and the original training code snapshot.
It excludes raw videos and training checkpoints. Older cache folders remain unchanged.

These are development comparisons: upstream T-WLA, RGB-world, and P1 training splits overlap the current test set.
Read the folder README before treating these results as a benchmark.

Code and experiment reports: [hannah/tactile_editor](https://github.com/kaichen-z/VLA-HAND/tree/hannah/tactile_editor/yuzhench_distill_editor/V4/tactile_editor).