TeLoGraF: Temporal Logic Planning via Graph-encoded Flow Matching

TL;DR

TeLoGraF tackles the challenge of planning under general Signal Temporal Logic (STL) specifications by introducing a graph-encoded flow matching framework. It encodes STL formulas as syntax trees using a Graph Neural Network and generates STL-satisfying trajectories via a conditional flow model, enabling fast, scalable planning across diverse dynamics. The authors compile a large dataset of over 200K STL specifications with demonstrations across five environments and demonstrate superior STL satisfaction rates and orders-of-magnitude faster inference than classical planners and several baselines. This work advances STL-driven robotics by enabling general STL conditioning, offering robust performance and a practical open-source resource for future research.

Abstract

Learning to solve complex tasks with signal temporal logic (STL) specifications is crucial to many real-world applications. However, most previous works only consider fixed or parametrized STL specifications due to the lack of a diverse STL dataset and encoders to effectively extract temporal logic information for downstream tasks. In this paper, we propose TeLoGraF, Temporal Logic Graph-encoded Flow, which utilizes Graph Neural Networks (GNN) encoder and flow-matching to learn solutions for general STL specifications. We identify four commonly used STL templates and collect a total of 200K specifications with paired demonstrations. We conduct extensive experiments in five simulation environments ranging from simple dynamical models in the 2D space to high-dimensional 7DoF Franka Panda robot arm and Ant quadruped navigation. Results show that our method outperforms other baselines in the STL satisfaction rate. Compared to classical STL planning algorithms, our approach is 10-100X faster in inference and can work on any system dynamics. Besides, we show our graph-encoding method's capability to solve complex STLs and robustness to out-distribution STL specifications. Code is available at https://github.com/mengyuest/TeLoGraF

Figures (49)

• Figure 1: Four STL templates used in this paper. Single-goal: reach one goal under time constraints while avoiding obstacles. Multi-goal: reach one of the valid subsets of the goals. Sequential: All the goals needed to be reached in a strict temporal order. Partial: Some goals must be reached first before reaching other goals (no global order is explicitly specified).
• Figure 2: System diagram for TeLoGraF. The STL specification is encoded as a graph structure data. The GNN extracted embedding and the state embedding are sent to the Temporal U-Net flow model to generate the planned trajectory (via ODE).
• Figure 3: Simulation benchmarks. In Linear and Dubins, a moving robot needs to reach circular regions and avoid circular obstacles. In PointMaze and AntMaze, the agent needs to reach/avoid square tiles in a maze. In Franka Panda, the robot arm needs to reach certain cubes on the table while not colliding with red balls.
• Figure 4: Main results. Our methods outperform both learning-based methods (CTG and LTLDoG) and classical methods (Grad, CEM). In four out of the five benchmarks, TeLoGraD achieves the highest solution quality. TeLoGraF (Fast) achieves the best trade-off between solution quality and efficiency over all benchmarks, especially being 123.6X faster than Grad and 60.7X faster than CEM with a higher satisfaction rate in the Franka Panda environment.
• Figure 5: Different encoder architecture comparisons. GNN and TreeLSTM works the best on the validation set.