Research map / Reinforcement learning / Visual and representation RL
Synthesized Policies for Transfer and Adaptation across Tasks and Environments
Hexiang Hu, Liyu Chen, Boqing Gong, Fei Sha
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 |
|---|---|---|---|
| compose policy from environment and task embeddings via shared basis (SynPo) instead of MLP on concatenated embeddings | imitation-learned agent · GridWorld unseen pairs | AvgSR: 66.1 ± 2.6% -> 82.1 ± 1.5% | Improved |
| model environment embeddings (SynPo) instead of ignoring environments (MTL) (ablation) | policy network · GridWorld unseen pairs | AvgSR: 41.5 ± 1.4% -> 82.1 ± 1.5% | Improved |
| reduce agent view window to 1x1 cell | MTL baseline · GridWorld unseen pairs | AvgSR: 41.5 -> 6.1 | Hurt |
| increase fraction of seen pairs beyond 40% of all pairs (ablation) | SynPo · GridWorld 10 envs x 10 tasks | AvgSR on unseen | No clear effect |
| fine-tune with PPO reinforcement learning after imitation | SynPo · GridWorld unseen pairs | AvgSR: 82.1 -> 96.16% | Improved |
| fine-tune with PPO instead of imitation only | MLP · GridWorld unseen pairs | AvgSR | Improved |
| transfer to unseen environments or tasks (setting 3, both new) instead of one new (setting 2) | SynPo · GridWorld cross pair (Q's env, P's task) | AvgSR: 50.5% -> 42.7% | Hurt |
| use SynPo instead of MLP when transferring to unseen environments/tasks | policy network · GridWorld transfer setting 2, cross pair (Q's env, P's task) | AvgSR: 13.8% -> 50.5% | Improved |
| use SynPo instead of MLP, cross pair (P's env, Q's task) | policy network · GridWorld transfer setting 2 | AvgSR: 20.7% -> 21.5% | No clear effect |
| use SynPo instead of MTL, ModuleNet, MLP | policy network · THOR unseen pairs | AvgSR: 33.3% -> 35.4% | Improved |
Closest papers by meaning
- Composing Task-Agnostic Policies with Deep Reinforcement Learning (ICLR 2020 · similarity 0.89)
- AMAGO-2: Breaking the Multi-Task Barrier in Meta-Reinforcement Learning with Transformers (NeurIPS 2024 · similarity 0.87)
- Task-Aware Virtual Training: Enhancing Generalization in Meta-Reinforcement Learning for Out-of-Distribution Tasks (ICML 2025 · similarity 0.86)
- Modeling Unseen Environments with Language-guided Composable Causal Components in Reinforcement Learning (ICLR 2025 · similarity 0.86)
- Meta-Reinforcement Learning with Adaptation from Human Feedback via Preference-Order-Preserving Task Embedding (ICML 2025 · similarity 0.86)
- Multi-Task Reinforcement Learning with Mixture of Orthogonal Experts (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.
