Research map / Reinforcement learning / Visual and representation RL
Zero-Shot Task Generalization with Multi-Task Deep 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 |
|---|---|---|---|
| add analogy-making objective to the parameterized skill (ablation) | Parameterized skill · Minecraft parameterized tasks, Independent unseen | average reward: -3.7 -> 0.3 | Helped |
| add analogy-making objective to the parameterized skill (ablation) | Parameterized skill · Minecraft parameterized tasks, Object-dependent unseen | average reward: -5.0 -> 0.3 | Helped |
| add analogy-making objective to the parameterized skill (ablation) | Parameterized skill · Minecraft parameterized tasks, Inter/Extrapolation unseen | average reward: -2.2 -> -1.7 | Helped |
| add analogy-making objective to the parameterized skill (ablation) | Parameterized skill · 2D grid-world parameterized tasks, Independent unseen | average reward: -1.88 -> 0.55 | Helped |
| add analogy-making objective to the parameterized skill (ablation) | Parameterized skill · 2D grid-world parameterized tasks, Object-dependent unseen | average reward: -3.23 -> 0.55 | Helped |
| learn when to update the subtask instead of updating every step (ablation) | Hierarchical meta controller · Minecraft instruction execution, test seen | average reward: -53.4 -> -30.3 | Helped |
| learn when to update the subtask instead of waiting for termination (ablation) | Hierarchical meta controller · Minecraft instruction execution, test unseen | average reward: -59.2 -> -38.0 | Helped |
| use a hierarchical skill-plus-meta-controller architecture rather than a flat controller (ablation) | Instruction execution · Minecraft instruction execution, train | average reward: -7.1 -> -3.1 | Helped |
| learn when to update the subtask rather than updating every time-step (ablation) | Hierarchical meta controller · 2D grid-world instruction execution, test unseen | reward: -17.23 -> -14.75 | Helped |
| use the learned-time-scale controller rather than the hand-designed Shortest Path policy | Instruction execution · 2D grid-world instruction execution, test | success rate: 99.4% -> 81.3% (seen); 99.4% -> 40.3% (unseen) | Hurt |
Closest papers by meaning
- Zero-Shot Instruction Following in RL via Structured LTL Representations (NeurIPS 2026 · similarity 0.89)
- Skill-aware Mutual Information Optimisation for Zero-shot Generalisation in Reinforcement Learning (NeurIPS 2024 · similarity 0.89)
- One Subgoal at a Time: Zero-Shot Generalization to Arbitrary Linear Temporal Logic Requirements in Multi-Task Reinforcement Learning (NeurIPS 2025 · similarity 0.88)
- Zero Shot Generalization of Vision-Based RL Without Data Augmentation (ICML 2025 · similarity 0.88)
- AMAGO-2: Breaking the Multi-Task Barrier in Meta-Reinforcement Learning with Transformers (NeurIPS 2024 · similarity 0.88)
- Task-Aware Virtual Training: Enhancing Generalization in Meta-Reinforcement Learning for Out-of-Distribution Tasks (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.