AI Paper Atlas

Research map / Reinforcement learning / Visual and representation RL

REACH: Explicit Recovery Behavior for Diffusion Policies

Zundong Ke, Junlin Chen, Jiayi Zhu, Kuanhao Xia, Boyi Zhao, Jiayuan Gu

CVPR 2026 · Reinforcement learning › Visual and representation RL · approach: Skill learning and diffusion policies

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add error detector with negative-prompt recovery to prompt-conditioned diffusion policy (ablation)REACH · Real robot Pick Cupsuccess rate: 66% -> 74%Improved
add error detector to prompt-conditioned diffusion policy (ablation)REACH · Real robot Stack Brickssuccess rate: 30% -> 42%Improved
add fixed base prompt conditioning without error detector (ablation)Diffusion Policy · Real robot Stack Brickssuccess rate: 32% -> 30%Hurt
add fixed base prompt conditioning without error detector (ablation)Diffusion Policy · Real robot Pick Cupsuccess rate: 58% -> 66%Improved
train error detector on PH data for policy trained on paired data (ablation)REACH · Robomimic Can (Paired)success rate: 64.1+-2.8 -> 74.1+-2.8Improved
increase negative guidance scale from 0 to 4.5 (ablation)REACH · Robomimic Can (PH detector)success rate: 74.1+-2.8 -> 85.0+-0Improved
excessively strong guidance scale (ablation)REACH · Robomimic / MimicGensuccess rateHurt
add negative guidance with larger scale (ablation)Diffusion Policy · PushT action steeringdeviation angle theta2: 41.59 -> 52.27 (seed 10000, scale 0.1 -> 0.3)Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy