Research map / Reinforcement learning / Visual and representation RL
Generalization to New Actions in Reinforcement Learning
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 |
|---|---|---|---|
| Use the hierarchical VAE instead of a non-hierarchical VAE for action representations | Proposed policy framework · Six evaluated tasks | policy performance | Helped |
| Use a learned task-conditioned utility function instead of continuous-output action selection | Proposed policy framework · Six evaluated tasks | zero-shot generalization performance | Helped |
| Use the learned utility policy instead of nearest-neighbor action selection | Proposed policy framework · Six evaluated tasks | test performance | Helped |
| Remove action-space subsampling during training (ablation) | Proposed policy · All evaluated tasks | test generalization performance | Hurt |
| Add action-space subsampling during training (ablation) | Proposed policy · All evaluated tasks | training performance | Hurt |
| Remove entropy regularization (ablation) | Proposed policy · Recommender System | test generalization performance | Hurt |
| Add entropy regularization (ablation) | Proposed policy · Evaluated tasks | training and test performance | Helped |
| Train from scratch on the new action set instead of using zero-shot transfer | Proposed adaptable policy · CREATE Push | environment steps to reach zero-shot performance: around 3 million | Hurt |
| Fine-tune a discrete policy after re-initializing its final layer instead of using zero-shot transfer | Discrete policy · CREATE Push | environment steps to reach zero-shot performance: over 1 million steps | Hurt |
| Continue training the pretrained adaptable policy on the new action set | Proposed adaptable policy · CREATE Push | convergence speed | Helped |
| Use an auxiliary policy conditioned on the selected discrete action rather than state encoding alone (ablation) | Proposed policy · Auxiliary policy alternative experiment | average success rate | No clear effect |
| Use video observations instead of state trajectories for CREATE action representations | Proposed HVAE and policy · CREATE | policy performance | No clear effect |
Closest papers by meaning
- Zero Shot Generalization of Vision-Based RL Without Data Augmentation (ICML 2025 · similarity 0.88)
- Foundation Policies with Hilbert Representations (ICML 2024 · similarity 0.88)
- Horizon Generalization in Reinforcement Learning (ICLR 2025 · similarity 0.88)
- Provable Zero-Shot Generalization in Offline Reinforcement Learning (ICML 2025 · similarity 0.87)
- How Ensembles of Distilled Policies Improve Generalisation in Reinforcement Learning (NeurIPS 2025 · similarity 0.87)
- Unlock the Cognitive Generalization of Deep Reinforcement Learning via Granular Ball Representation (ICML 2024 · 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.