Xun49/Event-aligned-robotwin

Event-Aligned — RoboTwin 2.0

Event-Aligned model for RoboTwin 2.0 bimanual manipulation, built on pretrained LingBot-VA base.

Code | Models | Project page

Event-aligned visual-action sampling (left) and DOMINO success rates under a matched training sequence budget (right).

Method

• Event-aligned supervision: preserve dense interaction evidence and sample transitions more sparsely, keeping training chunks within event boundaries.
• Execution validity prediction: identify valid action slots and skip redundant actions introduced by repeated-frame padding.

Inference uses observations and a task instruction. No event annotations are required during inference.

Download

Download the RoboTwin model from the model collection:

python -m pip install -U huggingface_hub
hf download Xun49/Event-aligned-robotwin --local-dir ./model/event-aligned_robotwin

Download the complete repository, including transformer/, vae/, text_encoder/, and tokenizer/. See the code repository for setup, inference, and evaluation.

Results reported in the paper

Method	Clean SR (%)	Randomized SR (%)	Average SR (%)
Event-Aligned	93.02	91.58	92.30

Evaluation covers 50 RoboTwin 2.0 tasks under clean and randomized conditions, with 100 episodes per task for each condition.

License and attribution

The model weights are released under the Apache License 2.0.

Built on the Robbyant Team's LingBot-VA. The companion code is also Apache-2.0 licensed. Retain applicable upstream licenses and attribution. Please also credit Causal World Modeling for Robot Control.

Citation

@misc{yang2026eventaligned,
  title         = {Event-Aligned Visual Action Reasoning for World Action Models},
  author        = {Yang, Xiaomeng and Wu, Yushu and Gao, Yi and Lei, Yuhao and Zhang, Xuan and Zhao, Pu and Wang, Yanzhi},
  year          = {2026},
  eprint        = {2610.09427},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2610.09427}
}