Research map / Graph learning / Knowledge and temporal graphs
TeLoGraF: Temporal Logic Planning via Graph-encoded Flow Matching
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 TeLoGraF, a GNN-encoded conditional flow-matching planner that generates trajectories satisfying general signal temporal logic (STL) specifications, trained on 200K STL specs with demonstrations across five simulated robot environments.
Problem
Planning trajectories that satisfy general STL specifications with one model, since prior learning methods handle only fixed or parametrized STLs and lack diverse STL datasets and effective STL encoders.
Approach
Encode the STL syntax tree as a directed graph with a GNN, and use the graph embedding to condition a flow-matching (linear flow) Temporal U-Net that generates trajectories; also build a dataset of 200K STLs from four templates with solver-generated demonstrations.
Key results
- TeLoGraF (Fast) is 123.6X faster than Grad and 60.7X faster than CEM on Franka Panda with a higher STL satisfaction rate.
- TeLoGraD achieves the highest STL satisfaction in four of five benchmarks, outperforming CTG, LTLDoG, Grad and CEM.
- GNN and TreeLSTM encoders generalize better than GRU and Transformer to unseen STL syntax and are more robust to augmented STL specifications.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| reduce sampling from 100 to 2 ODE steps (ablation) | TeLoGraF · Dubins training split | STL satisfaction rate: 0.97 -> 0.94 | Hurt |
| reduce sampling from 100 to 2 ODE steps (ablation) | TeLoGraF · Dubins validation split | STL satisfaction rate: 0.45 -> 0.34 | Hurt |
| reduce sampling steps to 10 (ablation) | TeLoGraF · Dubins | STL satisfaction rate | No clear effect |
| replace diffusion models with flow matching (ablation) | TeLoGraD · Dubins and Franka Panda | STL satisfaction rate | Hurt |
| use a GNN encoder instead of sequence-based GRU or Transformer encoding (ablation) | TeLoGraF · Unseen STL syntax | STL satisfaction rate | Helped |
| use graph-based encoding instead of sequence-based encoding (ablation) | TeLoGraF · Out-domain STL augmentations | STL satisfaction rate | Helped |
| use graph-based encoding instead of sequence-based encoding (ablation) | TeLoGraF · Sequential and Partial STL categories | STL satisfaction rate | Helped |
| use TeLoGraF (Fast) instead of TeLoGraF (ablation) | TeLoGraF · Dubins | runtime: 1/10 of the runtime | Helped |
| use TeLoGraF (Fast) instead of TeLoGraF (ablation) | TeLoGraF · Dubins | STL satisfaction rate | No clear effect |
| use TeLoGraF (Fast) instead of Grad | Franka Panda · Franka Panda | STL satisfaction rate | Helped |
Surprising findings
- Flow-matching TeLoGraF performs worse than diffusion-based TeLoGraD on Dubins and Franka Panda, attributed to nonlinear dynamics or forward kinematics.
- Even complex STL templates (multi-goal, sequential, partial) remain hard, with low satisfaction on Linear, Dubins and Franka Panda.
Limitations
- Data-driven, no soundness or completeness guarantees.
- Performance degrades on complex or heavily out-of-distribution STL syntax.
- Validation satisfaction for complex templates (multi-goal, sequential, partial) is low.
Experimental setup
Datasets: ZoneEnv Linear/Dubins demonstrations (generated) PointMaze demonstrations (generated via A* and PD controller) AntMaze demonstrations (generated via A* and RL goal-reaching policy) Franka Panda demonstrations (generated via gradient-based solver)
Benchmarks: Linear Dubins PointMaze AntMaze Franka Panda
Metrics: STL satisfaction rate runtime
Baselines: Grad Grad-lite CEM CTG LTLDoG GRU Transformer TreeLSTM TeLoGraD
Builds on: flow matching CTG LTLDoG GCN Temporal U-Net
Closest papers by meaning
- Zero-Shot Trajectory Planning for Signal Temporal Logic Tasks (NeurIPS 2025 · similarity 0.91)
- SAGAS: Semantic-Aware Graph-Assisted Stitching for Offline Temporal Logic Planning (NeurIPS 2026 · similarity 0.90)
- TGPO: Temporal Grounded Policy Optimization for Signal Temporal Logic Tasks (NeurIPS 2026 · similarity 0.87)
- Hierarchical Planning for Complex Tasks with Knowledge Graph-RAG and Symbolic Verification (ICML 2025 · similarity 0.86)
- Plan-Seq-Learn: Language Model Guided RL for Solving Long Horizon Robotics Tasks (ICLR 2024 · similarity 0.86)
- The Logical Expressiveness of Temporal GNNs via Two-Dimensional Product Logics (NeurIPS 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.