机械臂做到第10步就容易出错？一个 2B 模型靠动态调整注意力解决了 | IJCAI 2026
Ranking
Overall
73
Content
85
Popularity
46
Observed public metrics from 1 member.
Merged summary
TL;DR - S²-VLA is a 2B-parameter vision-language-action model that uses a learned belief state and dynamic attention gating to reduce cumulative errors in long-horizon robotic manipulation. It reports 96.4% success on LIBERO-Long while using about 7GB of inference memory, outperforming many 7B–8.5B models.
- A lightweight recurrent network derives a belief state from action history and joint-sensor feedback, learning task progress and execution quality without explicit stage labels.
- Its SSGAA module dynamically balances local visual attention, global intent attention, and action self-attention according to the current manipulation stage.
- Visual weighting rises for precise alignment, intent weighting rises during grasping and subtask transitions, and action self-attention dominates steady movement.
- The model reportedly runs at 80.8Hz, highlighting stage-adaptive computation as an efficient alternative to simply scaling VLA parameter counts.
Sources (1)
机械臂做到第10步就容易出错？一个 2B 模型靠动态调整注意力解决了 | IJCAI 2026
Public signals
Semantic Scholar citations 1 · Semantic Scholar influential citations 0
TL;DR - S²-VLA is a 2B-parameter vision-language-action model that uses a learned belief state and dynamic attention gating to reduce cumulative errors in long-horizon robotic manipulation. It reports 96.4% success on LIBERO-Long while using about 7GB of inference memory, outperforming many 7B–8.5B models.
- A lightweight recurrent network derives a belief state from action history and joint-sensor feedback, learning task progress and execution quality without explicit stage labels.
- Its SSGAA module dynamically balances local visual attention, global intent attention, and action self-attention according to the current manipulation stage.
- Visual weighting rises for precise alignment, intent weighting rises during grasping and subtask transitions, and action self-attention dominates steady movement.
- The model reportedly runs at 80.8Hz, highlighting stage-adaptive computation as an efficient alternative to simply scaling VLA parameter counts.