Research map / Reinforcement learning / Visual and representation RL
Dynamics Generalisation in Reinforcement Learning via Adaptive Context-Aware Policies
Michael Beukman, Devon Jarvis, Richard Klein, Steven James, Benjamin Rosman
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 |
|---|---|---|---|
| ignore context (Unaware) instead of giving the policy the context | SAC policy · ODE, train {-5,-1,1,5}, eval [-10,10] | average evaluation reward | Hurt |
| use Decision Adapter instead of Concat, cGate, FLAP | SAC policy · ODE 1D context | average evaluation reward | Improved |
| use Decision Adapter instead of baselines when extrapolating | SAC policy · ODE extrapolation contexts | average evaluation reward | Improved |
| use Decision Adapter instead of Concat on 2D context ODE | SAC policy · ODE 2D context | average evaluation reward | Improved |
| add context to the policy (Concat/cGate/Adapter) versus Unaware | SAC policy · CartPole | steps to reach reward above 400: 600 -> 100 | Improved |
| add distractor context variables (0 to 100) | Concat and cGate · CartPole | average evaluation reward | Hurt |
| add distractor context variables (0 to 100) | Decision Adapter · CartPole | average evaluation reward | No clear effect |
| sample non-fixed Gaussian distractors (train N(1,0.2), eval N(0,0.2)) | Decision Adapter vs Concat and cGate · CartPole | average evaluation reward | Improved |
| add distractor dimensions with different train and test values | Decision Adapter vs Concat and cGate · Mujoco Ant | average evaluation reward | Improved |
| use hidden width 8 or no hidden layers in the adapter instead of bottleneck 32 (ablation) | Decision Adapter · ODE | average evaluation reward | Hurt |
| use larger adapter hidden layer of 256 or multiple layers (ablation) | Decision Adapter · ODE | average evaluation reward | No clear effect |
| use hypernetwork chunking instead of non-chunking hypernetworks (ablation) | Decision Adapter · ODE | average evaluation reward | Improved |
Closest papers by meaning
- Zero-Shot Context Generalization in Reinforcement Learning from Few Training Contexts (NeurIPS 2025 · similarity 0.88)
- Towards Generalizable Reinforcement Learning via Causality-Guided Self-Adaptive Representations (ICLR 2025 · similarity 0.88)
- Behavior-agnostic Task Inference for Robust Offline In-context Reinforcement Learning (ICML 2025 · similarity 0.87)
- Behavioral Exploration: Learning to Explore via In-Context Adaptation (ICML 2025 · similarity 0.87)
- In-Context Reinforcement Learning for Variable Action Spaces (ICML 2024 · similarity 0.87)
- In-Context Reinforcement Learning From Suboptimal Historical Data (ICML 2025 · similarity 0.87)
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.
