Research map / Reinforcement learning / Visual and representation RL
ASPiRe: Adaptive Skill Priors for Reinforcement Learning
Mengda Xu, Manuela Veloso, Shuran Song
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 |
|---|---|---|---|
| Use adaptive weighted regularization from multiple primitive skill priors instead of one aggregated skill prior | ASPiRe · Point Maze, Ant Push, Ant Maze | learning efficiency and task success rate | Improved |
| Guide learning with a skill prior rather than train from scratch or behavior cloning with finetuning | Downstream tasks · Point Maze, Ant Push, Ant Maze | task success | Improved |
| Use composite skills as regularization rather than directly executing the composite as a policy (ablation) | ASPiRe · Point Maze, Ant Maze, and other evaluated tasks | learning performance | Improved |
| Use composite skill prior as a policy rather than as regularization (ablation) | ASPiRe · Point Maze and Ant Maze | learning performance | Hurt |
| Guide learning with only the avoid primitive (ablation) | Ant Maze · Ant Maze | task success | Hurt |
| Guide learning with only the navigation primitive (ablation) | Ant Maze · Ant Maze | task success | Hurt |
| Guide learning with only the push primitive (ablation) | Ant Maze · Ant Maze | task completion | Hurt |
| Use adaptive weights instead of uniform weights when some priors are irrelevant (ablation) | ASPiRe · Ant Maze without box pushing | learning performance | Improved |
| Use uniform weights when all skill priors are relevant (ablation) | ASPiRe · Ant Maze with all primitives relevant | learning performance | No clear effect |
| Vary critic-estimation sample size among M=1, 10, and 20 (ablation) | ASPiRe · Point Maze and Ant Maze | learning performance | No clear effect |
| Set target KL divergence too low (ablation) | ASPiRe · Point Maze and Ant Maze | learning performance: δ=6; δ=10 | Hurt |
| Set target KL divergence too high (ablation) | ASPiRe · Point Maze and Ant Maze | learning efficiency: δ=32; δ=50 | Hurt |
Closest papers by meaning
- Robust Policy Learning via Offline Skill Diffusion (AAAI 2024 · similarity 0.89)
- LLM-as-a-Tutor: Policy-Aware Prompt Adaptation for Non-Verifiable RL (NeurIPS 2026 · similarity 0.89)
- Efficient Skill Discovery via Regret-Aware Optimization (ICML 2025 · similarity 0.88)
- Adaptive Advantage-Guided Policy Regularization for Offline Reinforcement Learning (ICML 2024 · similarity 0.88)
- Leveraging Skills from Unlabeled Prior Data for Efficient Online Exploration (ICML 2025 · similarity 0.88)
- Dynamic Contrastive Skill Learning with State-Transition Based Skill Clustering and Dynamic Length Adjustment (ICLR 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.
