Rethinking World-Action Model for Compositional and In-Context Robotic Manipulation
ViGAR plans visual subgoals for robot manipulation, beating the best RoboTwin baseline by 12.86 points.
ViGAR is a hierarchical world-action framework that factorizes long-horizon manipulation into a visual subgoal planner and a subgoal executor sharing a pretrained world-model representation. A global goal image can induce different subtask decompositions in context without parameter updates. On RoboTwin Clean2Random it reaches 82.00% and 67.02% success under Clean and Random settings, exceeding the strongest baseline by 12.86 percentage points on average. Real-robot tests cover five compositional tasks and two in-context learning tasks.
- Splits manipulation into a visual subgoal planner and executor
- Both components share a pretrained world-model representation
- A goal image can change subtask behavior without weight updates
- RoboTwin success is 82.00% Clean and 67.02% Random
- Beats the strongest baseline by 12.86 points on average
From the original article
1 min read at huggingface.co
Long-horizon compositional manipulation has become increasingly important for real-world robot deployment, where a single task involves multiple coordinated subtasks. Existing world-action models (WAMs) jointly predict short-horizon visual futures and actions, but typically lack explicit subtask-level reasoning.
By Shukai Gong, Xuanran Zhai, Yintianrun Zhang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.