Coding Agents for Generalized Task and Motion Planning Problems

Matteo Merler, Bowen Li, Josh Roy, Yichao Liang, Qianwei Wang, Yixuan Huang, Tom Silver · 2026-09-24

This item is no longer on the current shelf. Its page remains available, and editorial corrections may be made. See what is featured now.

Coding agents synthesize reusable planners that outperform supplied planners in the tested simulated tasks.

What it does

A coding agent receives a task description, simulator access and a fixed synthesis budget. Its returned program is frozen before held-out testing, without an LLM at test time. The study evaluates 980 programs across 28 simulated environments. On the 16 environments with supplied planners, agent configurations achieve 56%–95% mean success versus 47% for those planners.

Why it matters

The experiment separates the cost of writing a reusable policy from executing it. Lower average per-instance computation as object counts grow does not establish lower total synthesis cost or successful physical deployment.

How it applies

Try the released setup in a simulator with disjoint test instances. Measure synthesis cost, execution time and failure cases together before considering deployment under perception uncertainty.

The short version

Generate a planner once; test its reuse separately.

My take

I would start with a small simulator experiment and a fixed budget, then compare the frozen program against an existing planner. Runtime savings would matter only alongside synthesis costs and unacceptable failures.

How it connects

• Compare reuse across held-out instances with repeated per-instance planning.
• Keep perception uncertainty outside the demonstrated scope until separately tested.

Bottom line

Reproduce the frozen-program protocol before adopting it. Primary paper.

Takeaways

• Programs are synthesized once and frozen before unseen-instance evaluation.
• The planner comparison covers 16 of the 28 simulated environments.
• The authors explicitly assume fully observed, object-centric states and simulator access.

Topics: task-and-motion planning, coding agents, program synthesis, LLM planning, generalization, scaling, ai-engineering