Research map / Reinforcement learning / Visual and representation RL
Hindsight Relabeling is All You Need for Reach-Avoid Learning
Kevin Li, Marinka Zitnik
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a hindsight relabeling approach for offline goal-conditioned RL on reach-avoid tasks, aiming to handle avoid regions specified flexibly at evaluation time.
Problem
Offline goal-conditioned RL for reach-avoid tasks, where existing methods encode avoid-region information into an augmented state space and cost function, preventing flexible specification of novel avoid regions at evaluation time.
Approach
Uses hindsight relabeling to handle reach-avoid objectives, so avoid-region information need not be baked into the augmented state or cost function; the abstract's full method description is truncated.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| encode goal and avoid-regions as prompt tokens (RADT) instead of augmented state with cost function, evaluated zero-shot on larger avoid boxes | RADT vs retrained AM-Lag · FetchReachObstacle, box width 0.24 | MNC: 0.457 (AM-Lag retrained) -> 0.409 (RADT zero-shot) | Improved |
| use RADT zero-shot on more avoid-regions than seen in training | RADT · UMazeObstacle, 5 avoid-regions | MNC (1e-2): 10.14 (AM-Lag retrained) -> 8.75 | Improved |
| RADT vs AM-Lag in-distribution on FetchReachObstacle | RADT · FetchReachObstacle, box width 0.16 | MNC: 0.011 (AM-Lag) vs 0.049 (RADT) | Hurt |
| RADT zero-shot with more avoid-regions in AMaze/HMaze where AM-Lag is in-distribution | RADT · HMazeObstacle, 7 avoid-regions | MNC (1e-2): 6.5 vs 6.03 (AM-Lag) | No clear effect |
| add avoid success indicator, AS, GS tokens, relabeling and attention boosting to prompt (ablation) | RADT · ablation suite (FetchReachObstacle / Maze) | MNC and SR | Improved |
| mix expert-policy trajectories into random training data (ablation) | RADT · OOD box sizes | MNC and SR | Mixed |
| mix expert-policy trajectories into random training data (ablation) | RADT · in-distribution box sizes | performance | No clear effect |
| use spherical (centroid+radius) avoid-region representation instead of box coordinates (ablation) | RADT · FetchReachObstacle | reach-avoid performance | No clear effect |
| zero-shot RADT up to 20 avoid-regions vs retrained AM-Lag | RADT · MazeObstacle, up to 20 avoid-regions | SR | Improved |
Closest papers by meaning
- PIPER: Primitive-Informed Preference-based Hierarchical Reinforcement Learning via Hindsight Relabeling (ICML 2024 · similarity 0.87)
- Stitching Sub-trajectories with Conditional Diffusion Model for Goal-Conditioned Offline RL (AAAI 2024 · similarity 0.87)
- Critic-Guided Decision Transformer for Offline Reinforcement Learning (AAAI 2024 · similarity 0.87)
- Flattening Hierarchies with Policy Bootstrapping (NeurIPS 2025 · similarity 0.87)
- Diffusion Subgoal Planning for Long-Horizon Offline Goal-Conditioned Reinforcement Learning (NeurIPS 2026 · similarity 0.87)
- Dual RL: Unification and New Methods for Reinforcement and Imitation Learning (ICLR 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.
