Research map / Reinforcement learning / Visual and representation RL
Learning Neuro-Symbolic Abstractions for Robot Planning and Learning
Read the paperCheck your idea against 108,967 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.
Doctoral-consortium style overview of a research program that learns neuro-symbolic state and action abstractions (critical-region-based Voronoi diagrams) to drive hierarchical robot planners, summarizing prior ATAM, HARP and SHARP results.
Problem
Hierarchical robot planners need hand-designed, environment-dependent state and action abstractions, while non-hierarchical planners fail on long-horizon tasks; the paper asks how to learn such hierarchical abstractions automatically for unseen environments and use them for efficient, reliable plann
Approach
Learn a critical-region predictor from randomly generated motion plans in training environments, use predicted critical regions to build a region-based Voronoi diagram that yields discrete abstract states and transition-based abstract actions, and plan hierarchically with interleaved symbolic search and low-level motion refinement.
Key results
- HARP with learned zero-shot state and action abstractions improves planning efficiency by roughly 10x over sampling-based and learning-based motion planners (as summarized from Shah and Srivastava 2022b).
- SHARP outperforms RRT, SAC and several HRL baselines across 14 stochastic settings, where most baselines fail to compute solutions (as summarized from Shah and Srivastava 2022a).
- ATAM uses a knapsack-based greedy refinement ordering so the robot begins executing actions much earlier than random selection (as summarized from Shah et al. 2020).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use hierarchical planning alongside learning | HARP · Multiple settings | efficiency: ∼10× | Helped |
| Prioritize likely outcomes for refinement with the greedy knapsack approach | ATAM · Combined task and motion planning | time until action execution | Helped |
| Use learned RBVD abstractions, options, and high-level A∗ planning | SHARP · 14 different settings | performance | Helped |
Limitations
- Relies on the critical-region predictor trained in a few training environments, with generalization to unseen environments as the central challenge.
- Hierarchical abstractions still require sound consistency with low-level motion planning.
Experimental setup
Datasets: randomly generated motion plans in training environments
Metrics: planning efficiency solution rate
Baselines: RRT SAC LLP PRM ATAM
Builds on: LLP ATAM HARP SHARP Beam search
Closest papers by meaning
- VisualPredicator: Learning Abstract World Models with Neuro-Symbolic Predicates for Robot Planning (ICLR 2025 · similarity 0.89)
- Hierarchical Planning and Learning for Robots in Stochastic Settings Using Zero-Shot Option Invention (AAAI 2024 · similarity 0.88)
- Hierarchical Planning for Complex Tasks with Knowledge Graph-RAG and Symbolic Verification (ICML 2025 · similarity 0.88)
- Reconciling Spatial and Temporal Abstractions for Goal Representation (ICLR 2024 · similarity 0.87)
- Learning Planning Abstractions from Language (ICLR 2024 · similarity 0.87)
- Efficient Robotic Policy Learning via Latent Space Backward Planning (ICML 2025 · similarity 0.86)
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.