Zhejiang Univ & TeleAI Propose GR-BC Using Graph Search for Robot Action Learning

1 reports archived1 independent sourcesupdated 11/2/2026, 18:19:00

Synthesis & Latest Updates

1 Sources Cross-Validated

Zhejiang University and TeleAI introduced GR-BC at CoRL 2026, a method that constructs state transition graphs and performs reverse search to extract effective action sequences from multiple trial-and-error robot trajectories containing invalid adjustments. This approach addresses the scarcity of perfect demonstrations in teleoperation data by using graph structures to evaluate action progress value, significantly enhancing the robustness and learning efficiency of Vision-Language-Action models in precise manipulation tasks.

LATEST/Zhejiang University and TeleAI introduced GR-BC at CoRL 2026, a method that constructs state transition graphs and performs reverse search to extract effective action sequences from multiple trial-and-error robot trajectories containing invalid adjustments. This approach addresses the scarcity of perfect demonstrations in teleoperation data by using graph structures to evaluate action progress value, significantly enhancing the robustness and learning efficiency of Vision-Language-Action models in precise manipulation tasks.

Hourly heat curve

2 fully observed hours

Fewer than 3 fully observed hours; trend not plotted yet.

Coverage timeline

Total 1 reports · Latest first

机器之心 (微信公众号)T2Nov 2, 18:19·78 pts
Zhejiang Univ & TeleAI Propose GR-BC Using Graph Search for Robot Action Learning
- GR-BC integrates scattered trial-and-error trajectories into a state transition graph, identifying optimal path segments across different attempts via reverse search.
- It quantifies the contribution of actions to task progress using graph structure, providing weighted guidance for behavior cloning training to correct policy bias.
- The paper and code are publicly available, addressing model performance bottlenecks caused by inconsistent quality of teleoperation data in fine-grained manipulation.
View Related Reports & Analysis Original Source