Research map / Probabilistic and scientific modeling / Dynamical systems
Zero-Shot Trajectory Planning for Signal Temporal Logic Tasks
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.
Proposes a hierarchical zero-shot STL planner that decomposes STL into time-aware progresses, searches timed waypoints using a learned time predictor, and stitches diffusion-generated trajectory segments (with SafeDiffuser constraints) trained on task-agnostic data.
Problem
Generating executable plans that satisfy Signal Temporal Logic (STL) specifications for systems with unknown dynamics, using only task-agnostic offline trajectory data and generalizing zero-shot to new STL tasks.
Approach
Decompose STL into reachability and invariance progresses with time variable constraints, allocate timed waypoints by DFS guided by a learned Gaussian time predictor, then generate and stitch short diffusion trajectory segments with constrained sampling (SafeDiffuser) to satisfy the global STL formula.
Key results
- On Maze2D, the proposed method achieves over 80% execution success rate across all U-Maze task types versus RGD's 32-80%, and at least 69% in Large.
- Planning time is over 10x faster than RGD (e.g., 0.86s vs 13.43s on U-Maze type 1).
- Cube success rate exceeds 84% and AntMaze exceeds 60% across nine STL templates each.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the proposed decomposition-allocation-generation planner instead of RGD | Maze2D U-Maze · Maze2D U-Maze Types 1–3 | execution success rate: 80.00 -> 97.33; 36.67 -> 92.00; 32.00 -> 91.33 | Helped |
| Use the proposed decomposition-allocation-generation planner instead of RGD | Maze2D Medium · Maze2D Medium Types 1–3 | execution success rate: 70.00 -> 94.67; 34.67 -> 89.33; 35.33 -> 83.33 | Helped |
| Use the proposed decomposition-allocation-generation planner instead of RGD | Maze2D Large · Maze2D Large Types 1–3 | execution success rate: 34.67 -> 92.00; 16.67 -> 81.33; 26.67 -> 79.33 | Helped |
| Use the proposed decomposition-allocation-generation planner instead of RGD | Maze2D U-Maze · Maze2D U-Maze Types 1–3 | robustness value: 0.1084$\pm$0.1132 -> 0.1938$\pm$0.0715; -0.2208$\pm$0.2826 -> 0.1504$\pm$0.0814; -0.1695$\pm$0.2521 -> 0.1354$\pm$0.0745 | Helped |
| Use the proposed decomposition-allocation-generation planner instead of RGD | Maze2D Medium · Maze2D Medium Types 1–3 | robustness value: 0.0885$\pm$0.2768 -> 0.3205$\pm$0.0957; -0.2277$\pm$0.3089 -> 0.2013$\pm$0.1950; -0.2393$\pm$0.3130 -> 0.1761$\pm$0.2060 | Helped |
| Use the proposed decomposition-allocation-generation planner instead of RGD | Maze2D Large · Maze2D Large Types 1–3 | robustness value: -0.1927$\pm$0.3198 -> 0.3180$\pm$0.1228; -0.3926$\pm$0.2284 -> 0.1672$\pm$0.2411; -0.2886$\pm$0.2934 -> 0.1639$\pm$0.228 | Helped |
| Replace RGD with the proposed planner | Maze2D · Maze2D U-Maze, Medium, and Large; Types 1–3 | total planning time: >10× faster | Helped |
| Generate plans with the proposed framework | Cube and AntMaze · OGBench Cube and AntMaze | execution success rate: over 84%; over 60% | Helped |
| Use progress allocation to identify feasible solutions | Proposed planner · Custom environment, verified feasible cases | success rate: exceeding 80% | Helped |
Limitations
- Algorithm is incomplete and conservative due to disjunction elimination, heuristic greedy allocation, and data-driven synthesis.
- Trajectory generation (SafeDiffuser QP) is the main runtime bottleneck.
- Experiments focus on navigation-style STL tasks; time predictor may generalize poorly with sparse data.
Experimental setup
Datasets: D4RL OGBench
Benchmarks: Maze2D U-Maze Maze2D Medium Maze2D Large OGBench Cube OGBench AntMaze
Metrics: execution success rate average robustness value total planning time trajectory generation time
Baselines: Robustness Guided Diffuser (RGD) sound and complete optimization-based STL planner
Builds on: SafeDiffuser Robustness Guided Diffuser stlpy
Closest papers by meaning
- TeLoGraF: Temporal Logic Planning via Graph-encoded Flow Matching (ICML 2025 · similarity 0.91)
- SAGAS: Semantic-Aware Graph-Assisted Stitching for Offline Temporal Logic Planning (NeurIPS 2026 · similarity 0.89)
- STL: Still Tricky Logic (for System Validation, Even When Showing Your Work) (NeurIPS 2024 · similarity 0.87)
- TGPO: Temporal Grounded Policy Optimization for Signal Temporal Logic Tasks (NeurIPS 2026 · similarity 0.87)
- Plan-Seq-Learn: Language Model Guided RL for Solving Long Horizon Robotics Tasks (ICLR 2024 · similarity 0.87)
- Hierarchical Planning for Complex Tasks with Knowledge Graph-RAG and Symbolic Verification (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.