Research map / Reinforcement learning / Visual and representation RL
Modular Multitask Reinforcement Learning with Policy Sketches
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 sketch-guided modular subpolicies shared across tasks instead of a shared task-conditioned policy | Multitask policies · Crafting multitask | task completion rate: .49 -> .89 | Helped |
| Use sketch-guided modular subpolicies shared across tasks instead of independent task policies | Multitask policies · Crafting multitask | task completion rate: .44 -> .89 | Helped |
| Use sketch-guided modular subpolicies instead of unsupervised option discovery | Multitask policies · Crafting multitask | task completion rate: .47 -> .89 | Helped |
| Provide a sketch and compose its learned modular subpolicies rather than condition a joint policy directly on the sketch | Held-out task policy · Crafting zero-shot | task completion rate: .01 -> .77 | Helped |
| Use reusable modular subpolicies as the basis for adaptation instead of independent task-specific learning | Held-out task policy · Crafting adaptation | task completion rate: .01 -> .76 | Helped |
| Add state dependence to a task-specific scalar critic (ablation) | Modular policy training · Crafting | convergence performance | Helped |
| Add task dependence to a state-dependent critic (ablation) | Modular policy training · Crafting | convergence performance | Helped |
| Use a critic dependent on both state and task rather than either dependence alone (ablation) | Modular policy training · Crafting | convergence performance | Helped |
| Remove both state and task dependence from the critic, leaving a constant baseline (ablation) | Modular policy training · Crafting | overall performance | Hurt |
| Add a curriculum that starts with short sketches and progresses to longer ones (ablation) | Modular policy training · Crafting | convergence rate | Helped |
| Sample tasks in inverse proportion to their accumulated reward (ablation) | Modular policy training · Crafting | convergence rate | Helped |
| Combine sketch-length progression with reward-weighted task sampling (ablation) | Modular policy training · Crafting | convergence rate | Helped |
Closest papers by meaning
- Multi-Task Reinforcement Learning with Soft Modularization (NeurIPS 2020 · similarity 0.89)
- Representation Learning Enables Scalable Multitask Deep Reinforcement Learning (NeurIPS 2026 · similarity 0.89)
- Learning Novel Policies For Tasks (ICML 2019 · similarity 0.89)
- Reward-Respecting Subtasks for Model-Based Reinforcement Learning (Abstract Reprint) (AAAI 2024 · similarity 0.88)
- Code as Reward: Empowering Reinforcement Learning with VLMs (ICML 2024 · similarity 0.87)
- Contextual Pre-planning on Reward Machine Abstractions for Enhanced Transfer in Deep Reinforcement Learning (AAAI 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.