Wjjjh/lingbot-va-libero-object

LingBot-VA LIBERO-Object

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

The checkpoint predicts robot actions jointly with future visual observations. It is intended for research and evaluation in the LIBERO simulator.

Checkpoint and result

Repository path	Training step	LIBERO-Object result
transformer/	5,000	497 / 500 (99.4%)

The result is one completed evaluation run over all 10 LIBERO-Object tasks with 50 episodes per task. The run used the runtime-default seed, which was not explicitly recorded. It should therefore be treated as a single-run result, not a multi-seed estimate.

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

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

Training details

Setting	Value
Initialization	robbyant/lingbot-va-base
Dataset	Wjjjh/libero_lingbot_va, libero_object subset
Dataset size	500 episodes, 74,507 action rows
Optimizer	AdamW
Learning rate	1e-5
Betas	(0.9, 0.95)
Weight decay	0.1
Warmup	10 optimizer steps
Precision	bfloat16
Distributed strategy	FSDP
Hardware	4 × NVIDIA H100 80 GB
Per-GPU batch size	1
Gradient accumulation	8
Effective global batch size	32
Total optimizer steps	5,000
Checkpoint interval	200 steps

Training followed the LingBot-VA LIBERO post-training pipeline, with a suite-specific action normalization computed from all LIBERO-Object 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": [
    -0.543749988079071,
    -0.8758928775787354,
    -0.9375,
    -0.06964285671710968,
    -0.11785714328289032,
    -0.16285714507102966,
    -1.0
  ],
  "q99": [
    0.84375,
    0.84375,
    0.9375,
    0.0803571417927742,
    0.14785714447498322,
    0.0867857113480568,
    1.0
  ]
}

The reported evaluation used:

• LIBERO-Object, 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
• LIBERO commit 8f1084e3132a39270c3a13ebe37270a43ece2a01

Limitations

• The published result is currently based on one run with an unrecorded runtime-default seed.
• 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}
}