Research map / Reinforcement learning / Visual and representation RL
Self-Supervised Policy Adaptation during Deployment
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Add online test-time inverse-dynamics adaptation to SAC+IDM (ablation) | SAC+IDM · DMControl randomized colors | episodic return: 406\scriptstyle{\pm29} -> 468\scriptstyle{\pm47} | Helped |
| Add online test-time inverse-dynamics adaptation to SAC+IDM | SAC+IDM · DMControl video backgrounds | episodic return: 7 out of 8 tasks | Mixed |
| Add online test-time inverse-dynamics adaptation to SAC+IDM (ablation) | SAC+IDM · Real robot push, cloth | success rate: 64\% -> 88\% | Helped |
| Add online test-time inverse-dynamics adaptation to SAC+IDM (ablation) | SAC+IDM · Simulated robot push, all dynamics changes | success rate: 48\% -> 76\% | Helped |
| Use rotation prediction rather than inverse dynamics for PAD (ablation) | A2C navigation policy · CRLMaze | episodic return | Mixed |
| Replace inverse dynamics with CURL and retain test-time adaptation (ablation) | SAC policy · DMControl randomized colors | episodic return: 445\scriptstyle{\pm99} -> 495\scriptstyle{\pm70}; 691\scriptstyle{\pm12} -> 588\scriptstyle{\pm22} | Mixed |
| Replace inverse dynamics with rotation prediction and retain test-time adaptation (ablation) | SAC policy · DMControl randomized colors | episodic return: 335\scriptstyle{\pm7} -> 330\scriptstyle{\pm30}; 283\scriptstyle{\pm68} -> 230\scriptstyle{\pm53} | Mixed |
| Discard policy updates after each step instead of carrying adaptation forward online (ablation) | SAC+IDM (PAD) · DMControl randomized colors | episodic return: 468\scriptstyle{\pm47} -> 441\scriptstyle{\pm16}; 797\scriptstyle{\pm46} -> 727\scriptstyle{\pm21} | Mixed |
| Keep the self-supervised prediction head fixed during deployment adaptation (ablation) | SAC+IDM (PAD) · DMControl randomized colors | episodic return: 452\scriptstyle{\pm38} vs 468\scriptstyle{\pm47}; 802\scriptstyle{\pm41} vs 797\scriptstyle{\pm46} | No clear effect |
| Continue self-supervised adaptation in the unchanged training environment (ablation) | SAC+IDM (PAD) · DMControl training environments | episodic return: 911\scriptstyle{\pm24} -> 895\scriptstyle{\pm28}; 928\scriptstyle{\pm45} -> 927\scriptstyle{\pm45} | No clear effect |
| Use PAD with inverse dynamics rather than no test-time adaptation (ablation) | SAC+IDM · DMControl randomized colors, Cartpole balance | episodic return: 835\scriptstyle{\pm40} -> 848\scriptstyle{\pm29} | Mixed |
Closest papers by meaning
- Learning a subspace of policies for online adaptation in Reinforcement Learning (ICLR 2022 · similarity 0.89)
- Efficient Test-Time Adaptation For Robot Policies (NeurIPS 2026 · similarity 0.87)
- Foundation Policies with Hilbert Representations (ICML 2024 · similarity 0.87)
- Learning to Reuse Policies in State Evolvable Environments (ICML 2025 · similarity 0.87)
- How Ensembles of Distilled Policies Improve Generalisation in Reinforcement Learning (NeurIPS 2025 · similarity 0.86)
- Stabilizing Contrastive RL: Techniques for Robotic Goal Reaching from Offline Data (ICLR 2024 · 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.