LingBot-VA LIBERO-Spatial

This repository contains community post-trained transformer checkpoints for LingBot-VA on the LIBERO-Spatial suite. It is not an official Robbyant model release.

The checkpoints predict robot actions jointly with future visual observations. They are intended for research and evaluation in the LIBERO simulator.

Checkpoints

This checkpoint was retrained from robbyant/lingbot-va-base after replacing an earlier incompatible LIBERO-Spatial dataset export. Each evaluation covers all 10 tasks with 50 episodes per task.

The population standard deviation across the three seed-level success rates is 0.81 percentage points.

For reference, the official LingBot-VA repository reports 98.5% ± 0.3% for LIBERO-Spatial. This community checkpoint's three-seed mean is 3.17 percentage points lower, so it does not fully match the reported official result. The two figures should not be treated as strictly identical experiments unless every training and evaluation detail is matched.

Model contents

This repository provides the post-trained transformer only. The VAE, tokenizer, and text encoder must be taken from robbyant/lingbot-va-base.

The checkpoint config is saved with:

{"attn_mode": "flex"}

flex is the training setting. For inference or evaluation, copy the checkpoint into a runtime model directory and change attn_mode to torch or flashattn, following the official LingBot-VA instructions. Do not overwrite the archived training checkpoint merely to switch inference backends.

Download

Download the final 5,000-step transformer:

hf download Wjjjh/lingbot-va-libero-spatial \
  --include "transformer/*" \
  --local-dir lingbot-va-libero-spatial-step5000

Training details

Training followed the LingBot-VA LIBERO post-training pipeline, with a suite-specific action normalization computed from all LIBERO-Spatial training actions.

Action and evaluation configuration

The model uses the first seven channels of the 30-channel action layout:

used_action_channel_ids = [0, 1, 2, 3, 4, 5, 6]
action_snr_shift = 0.05

The action quantiles used for both training and evaluation are:

{
  "q01": [
    -0.7607142857142858,
    -0.6562499999999999,
    -0.9375,
    -0.10821428571428572,
    -0.20464285714285715,
    -0.18642857142857147,
    -1.0
  ],
  "q99": [
    0.9375,
    0.8732142857142857,
    0.9348214285714286,
    0.10500000000000001,
    0.17571428571428574,
    0.14357142857142857,
    1.0
  ]
}

The reported 5,000-step evaluations used:

• LIBERO-Spatial, 10 tasks
• 50 episodes per task (500 total)
• maximum 800 environment steps per episode
• 128 × 128 agent-view and eye-in-hand observations
• 4 parallel GPU shards
• explicit seeds 2026072501, 2026072502, and 2026072503
• LIBERO commit 8f1084e3132a39270c3a13ebe37270a43ece2a01

Limitations

• The reported rates are simulator- and seed-dependent estimates from three 500-episode runs, not guarantees of real-world performance.
• Results are simulator- and configuration-dependent.
• This model has not been validated for safe real-world robot deployment.
• These files contain model weights only and do not contain optimizer state, so they are not a complete resumable training checkpoint.

Acknowledgements and citation

This checkpoint is derived from the official LingBot-VA base model and training code. Please cite the original LingBot-VA work when using this model:

@article{lingbot-va2026,
  title={Causal World Modeling for Robot Control},
  author={Li, Lin and Zhang, Qihang and Luo, Yiming and Yang, Shuai and Wang, Ruilin and Han, Fei and Yu, Mingrui and Gao, Zelin and Xue, Nan and Zhu, Xing and Shen, Yujun and Xu, Yinghao},
  journal={arXiv preprint arXiv:2601.21998},
  year={2026}
}