The paper investigates whether can automate the synthesis of programs for generalized task and motion planning (TAMP) problems, which typically require substantial domain-specific engineering. The authors evaluated three coding agent configurations (Claude Code with Opus 5, Codex with GPT-5.6 Sol, and Codex with GPT-6 Astra) on 28 simulated environments from KinDER and PDDLStream. The agents were given a task description and simulator access, allowed to interact with the environment to develop a program within a fixed budget, and then the frozen program was tested on unseen instances. Across 98,000 evaluation episodes, the agents' programs achieved mean success rates between 56% and 95%, outperforming hand-engineered planners which averaged 47% on the 16 environments where planners were available. The agents' programs also maintained higher success as object counts grew, using an order of magnitude less computation per instance than the planners. This is interesting because it suggests that coding agents can reduce the need for TAMP-specific engineering while achieving better performance and efficiency.
Coding agents synthesize robot planners that generalize to unseen instances
Programs developed through simulator interaction surpassed hand-engineered planners while requiring substantially less computation at evaluation time.
Top University
Matteo Merler · Bowen Li · Josh Roy · Yichao Liang · Qianwei Wang · Yixuan Huang · +1 more
Fondazione Bruno Kessler · Carnegie Mellon University · Princeton University · University of Cambridge
Research Digest··2 min read
In one line
Coding agents outperform hand-engineered planners on generalized task and motion planning problems by synthesizing reusable programs.
Before you read this
Three ideas this paper builds oncoding agentAn AI-powered software system that autonomously writes, modifies, or extends source code based on high-level instructions.158 papers →amortized planningA learned model generates plans directly for new instances, replacing expensive per-instance search with a fast, generalizable forward pass.9 papers →code as actionA paradigm where actions are expressed as executable code, such as method calls or program synthesis, rather than direct manipulation.29 papers →
Read it as
The authors gave coding agents task descriptions and simulator access, then asked them to develop reusable programs for 28 simulated task-and-motion planning environments.
The authors evaluated Claude Code with Opus 5 and Codex with GPT-5.6 Sol or GPT-6 Astra on 28 environments drawn from KinDER and PDDLStream.
How it works
before and afterThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.