Differentiable SpaTiaL: Symbolic Learning and Reasoning with Geometric Temporal Logic for Manipulation Tasks

A team of researchers including Licheng Luo, Kaier Liang, Cristian-Ioan Vasile, and Mingyu Cai has introduced Differentiable SpaTiaL, a groundbreaking toolbox that makes spatio-temporal logic fully differentiable for the first time. The system addresses a critical limitation in robotics: while Spatio-Temporal Logic (SpaTiaL) provides a powerful framework for specifying complex manipulation tasks with geometric and temporal constraints, traditional implementations rely on non-differentiable geometric operations that break computational graphs and prevent gradient-based optimization. Differentiable SpaTiaL overcomes this by constructing smooth, autograd-compatible geometric primitives directly over polygonal sets, enabling exact gradient propagation through spatial operations.

By analytically deriving differentiable relaxations of key spatial predicates—including signed distance, intersection, containment, and directional relations—the toolbox creates an end-to-end differentiable mapping from high-level semantic specifications to low-level geometric configurations. This eliminates the need for external discrete solvers that previously disrupted gradient flow. The fully differentiable formulation unlocks two transformative capabilities: massively parallel trajectory optimization under rigorous spatio-temporal constraints, and direct learning of spatial logic parameters from demonstrations via backpropagation. Experimental results demonstrate the system's effectiveness and scalability for complex manipulation tasks in cluttered environments where robots must satisfy coupled geometric and temporal constraints.

The technical breakthrough lies in the toolbox's ability to maintain differentiability throughout the entire pipeline, from symbolic logic specifications to geometric configurations. This represents a significant advancement over existing approaches where differentiable temporal logics focused only on robot internal states while neglecting object-environment interactions, and spatial logic methods relied on discrete geometry engines. The arXiv preprint (2604.02643) details how Differentiable SpaTiaL enables robots to learn manipulation tasks more efficiently by allowing gradient-based optimization of both the trajectory and the spatial logic parameters simultaneously, potentially accelerating robot learning from demonstrations and improving performance in real-world manipulation scenarios.