AI Paper Atlas

Research map / Reinforcement learning / Visual and representation RL

For Pre-Trained Vision Models in Motor Control, Not All Policy Learning Methods are Created Equal

Yingdong Hu, Renhao Wang, Li Erran Li, Yang Gao

ICML 2023 · Reinforcement learning › Visual and representation RL · approach: Vision-based robot RL · 2 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 reward function (VRF) instead of RL for evaluationpre-trained vision models for motor control · Meta-World, Robosuite, Franka-Kitchenconsistency of model rankings across runs and environmentsImproved
Use RL (DrQ-v2) as downstream policy learning methodpre-trained vision models · Meta-World vs Robosuitemodel ranking consistencyHurt
Increase seeds from 3 to 40+ runs for RL comparisons (ablation)RL evaluation of vision models · Meta-World (5 tasks)stability of probability of improvementImproved
Use linear probing loss as a proxy for BC performance (lower loss)pre-trained vision models with BC · Franka-Kitchen (5 tasks)correlation with BC success rate: Pearson r = -0.78Improved
Use ImageNet k-NN accuracy as predictor of VRF performanceImageNet-pretrained vision models with VRF · all environmentscorrelation with VRF performance: r = 0.91Improved
Use MAE pre-trained features directly (masked image modeling) as frozen representations for VRFMAE with VRF · all environmentsIQM success rateHurt
Use joint embedding architectures (contrastive/self-distillation) rather than masked image modelingpre-trained vision models with VRF · all environmentsVRF success rateImproved
Use iBOT (MIM plus self-distillation) over other ViTsViT-B/16 models with BC · BC across environmentsBC performanceImproved
Use ResNet-50 backbone instead of ViT-B/16pre-trained vision models with BC · BC across environmentsBC performanceImproved

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