LingBot-VA overview
LingBot-VA is a causal video-action world model for robot control. It represents visual observations and actions in an interleaved autoregressive sequence and supports both visual dynamics prediction and action inference.
Released models
| Checkpoint | Use |
|---|---|
| lingbot-va-base | Base checkpoint for post-training |
| lingbot-va-posttrain-robotwin | Post-trained checkpoint for RoboTwin |
| lingbot-va-posttrain-libero-long | Post-trained checkpoint for LIBERO-Long |
Post-training datasets are available for RoboTwin and LIBERO-Long.
Runtime modes
The model configuration uses different attn_mode values for training and inference:
- Training:
flex - Inference or evaluation:
torchorflashattn
Set the value in the downloaded model's transformer/config.json before running the corresponding workflow.