Coding Agents for Generalized Task and Motion Planning Problems

Problem

Generalized Task and Motion Planning (TAMP) addresses the challenge of integrating discrete decision-making with geometric, kinematic, and dynamic constraints. This paper is a preprint and contributes to the literature by proposing a novel approach to TAMP that enhances generalization across various instances.

Method

The authors introduce coding agents that synthesize programs capable of generalizing across different task and motion planning scenarios. The architecture allows agents to autonomously determine their interaction strategies with the environment while developing a program within a predefined synthesis budget. The evaluation is conducted on 28 simulated environments derived from KinDER and PDDLStream. The training process involved generating 980 programs, which were then evaluated on 100 held-out instances, resulting in a total of 98,000 evaluation episodes. The evaluation method involved freezing the programs and testing them on unseen instances to assess their generalization capabilities.

Results

The coding agents achieved a mean success rate of 95% across 16 environments, significantly outperforming hand-engineered planners, which recorded a success rate of only 47%. Additionally, the programs generated by the agents demonstrated superior computational efficiency, maintaining a higher success rate while requiring an order of magnitude less computation per instance compared to traditional methods.

Limitations

The authors do not report any limitations in their work, and no obvious limitations are identified in the available text.

Why it matters

The implications of this research are significant for downstream applications in robotics and AI, as it provides a framework for developing more efficient and adaptable planning systems. The ability to generalize across diverse environments can lead to advancements in autonomous systems, enhancing their performance in real-world scenarios.