Research map / Reinforcement learning / Visual and representation RL
End-to-End Training of Deep Visuomotor Policies
Sergey Levine, Chelsea Finn, Trevor Darrell, Pieter Abbeel
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 changed | Where | Measured | Effect |
|---|---|---|---|
| replace softmax + fully connected layer with spatial softmax followed by expectation (feature points) (ablation) | Spatial softmax CNN · PR2 target pose estimation | test error (cm): 2.59 -> 1.30 | Improved |
| replace plain fully connected layer with spatial softmax + feature points (ablation) | Spatial softmax CNN · PR2 target pose estimation | test error (cm): 4.75 -> 1.30 | Improved |
| replace max-pooling + fully connected with spatial softmax + feature points (ablation) | Spatial softmax CNN · PR2 target pose estimation | test error (cm): 3.71 -> 1.30 | Improved |
| add max-pooling to the fully connected network | Fully connected CNN · PR2 target pose estimation | test error (cm): 4.75 -> 3.71 | Improved |
| increase training perturbation from 0 cm to 2 cm | Linear-Gaussian controller · PR2 lego block task at 2 cm test perturbation | successful trials out of 5: 3/5 -> 5/5 | Improved |
| increase training perturbation from 0 cm to 1 cm | Linear-Gaussian controller · PR2 ring on peg task at 2 cm test perturbation | successful trials out of 5: 0/5 -> 3/5 | Improved |
| increase training perturbation from 1 cm to 2 cm | Linear-Gaussian controller · PR2 ring on peg task at 2 cm test perturbation | successful trials out of 5: 3/5 -> 3/5 | No clear effect |
| use learned linear-Gaussian controller instead of kinematic baseline | Linear-Gaussian controller · PR2 lego block task at 1 cm test perturbation | successful trials out of 5: 0/5 -> 5/5 | Improved |
| train vision layers jointly end-to-end instead of pretraining for pose prediction (ablation) | Deep visuomotor policy · PR2 coat hanger, cube, hammer and bottle tasks | success rate | Improved |
| feed pretrained pose-estimate features (rather than predicted pose) to the control layers (ablation) | Deep visuomotor policy · PR2 visuomotor tasks | success rate | Improved |
| add Gaussian mixture model prior to trajectory optimization (ablation) | Guided policy search with linear-Gaussian controllers · simulated peg insertion, octopus arm, swimming | samples required | Improved |
Closest papers by meaning
- Neural Dynamic Policies for End-to-End Sensorimotor Learning (NeurIPS 2020 · similarity 0.90)
- CARP: Visuomotor Policy Learning via Coarse-to-Fine Autoregressive Prediction (ICCV 2025 · similarity 0.90)
- Latent Policy Barrier: Learning Robust Visuomotor Policies by Staying In-Distribution (NeurIPS 2025 · similarity 0.88)
- DynaMo: In-Domain Dynamics Pretraining for Visuo-Motor Control (NeurIPS 2024 · similarity 0.86)
- Scaling Proprioceptive-Visual Learning with Heterogeneous Pre-trained Transformers (NeurIPS 2024 · similarity 0.86)
- FreqPolicy: Frequency Autoregressive Visuomotor Policy with Continuous Tokens (NeurIPS 2025 · similarity 0.86)
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.
