Create

Sign in to ReadmeX

or

How we handle your data: Privacy Policy

ReadmeX

Your AI reading companion

What would you like to catch up on?

Pick a briefing, or ask your own question.

Your community and people briefings stay personal to you.

Story

GR-BC reuses useful actions from imperfect robot demonstrations

AI summary

Zhejiang University, TeleAI, and Gamma Intelligence proposed Graph-Reweighted Behavior Cloning (GR-BC), which links multiple robot attempts in a graph and uses reverse Dijkstra search to estimate progress toward successful outcomes. Action chunks receive positive training weights when they show sufficient net progress and limited retreat, allowing useful local behavior to be retained even in unsuccessful trajectories. With X-VLA as the backbone, the team reported final success rates of 98%, 93%, 89%, and 83% across four real-robot manipulation tasks, averaging about 91%.

In-depth insight

The method shifts policy learning from judging entire trajectories to identifying valuable local actions across connected attempts, and the paper and code are publicly available.

机器之心 2026-11-02 18:19 辽宁

有望帮助策略学习偏差纠正与受挫恢复,为更鲁棒的机器人操作提供支持。

用螺丝刀拧紧螺丝,听起来并不复杂。

但真正动手时,往往还要挪一点、转一点、退出来,再重新对齐。换成遥操作机器人采集数据,这个过程只会更考验耐心。即使是 “大国工匠” 般熟练的数采员,面对精细度要求极高的任务,也未必能一次交出从头到尾都足够丝滑的示范。一条最终成功的轨迹里,可能包含多次无效调整;一条最终因超时或碰撞而失败的尝试中,却也可能有几步对齐做得很好。

#Graph-Reweighted Behavior Cloning (GR-BC)#Zhejiang University#TeleAI#Gamma Intelligence#X-VLA#MCR#SmolVLA#π₀.₅

Original · 6 min readRead the original
Group chat

No comments yet. Start the conversation.