AI Paper Atlas

Research map / Reinforcement learning / Visual and representation RL

Self-Improving Loops for Visual Robotic Planning

Calvin Luo, Zilai Zeng, Mingxi Jia, Yilun Du, Chen Sun

ICLR 2026 · Reinforcement learning › Visual and representation RL · approach: Goal-conditioned policy learning · 8 citations (OpenAlex)

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
use visual planning (SILVR) instead of behavior-cloning improvement loopSILVR vs BCIL · MetaWorld 12 unseen taskssuccess rate (%): BCIL 23.2 -> SILVR 44.2 (iter 4)Improved
self-improve with RL finetuning (DSRL) instead of SILVRDSRL · MetaWorld 12 unseen taskssuccess rate (%): DSRL 9.4 -> 7.7Hurt
replace ground-truth success filter with VLM filter (ablation)SILVR · MetaWorld 12 unseen taskssuccess rate (%): 44.2 -> 38.4 (iter 4)Mixed
finetune both video model and IDM during self-improvement (ablation)SILVR · MetaWorldsuccess rate (%): 26.8 / 29.8 -> 44.2 (iter 4)Improved
distill final video planner into diffusion policySILVR · MetaWorldsuccess rate (%): 44.2 -> 49.2Improved
add internet video prior (AnimateDiff) via IPA (ablation)SILVR · Panda arm real-world taskssuccess rate (%)Improved
add internet video prior via IPA (ablation)SILVR · MetaWorldsuccess rate (%)No clear effect
skip data filtering (ablation)SILVR · MetaWorldsuccess rate (%)Hurt
skip data filtering (with internet prior) (ablation)SILVR · Panda armsuccess rate (%)No clear effect
initialize with suboptimal (70% random action) demonstrations (ablation)SILVR · MetaWorld 12 unseen taskssuccess rate (%): 44.2 (expert) vs 41.0 (suboptimal) at iter 4Mixed
train DSRL with 60000 instead of 150 updates (ablation)DSRL · MetaWorldsuccess rate (%): 8.3 vs 7.7 (iter 4)No clear effect
run more SILVR iterations beyond 5 (ablation)SILVR · MetaWorldsuccess rate (%): 51.1 (iter 5) -> 53.1 (iter 9)No clear effect

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