Robotics Research Hub
Coding Agents for Generalized Task and Motion Planning Problems

Coding Agents for Generalized Task and Motion Planning Problems

Matteo Merler, Bowen Li, Josh Roy, Yichao Liang, Qianwei Wang, Yixuan Huang, Tom Silver · N/A · 2026

vla planning benchmark
PDF arXiv

Framework

N/A

License

N/A

Stars

N/A

Summary

Task and motion planning (TAMP) problems remain difficult even with full observability and object-centric states because discrete decisions are tightly coupled to geometric, kinematic, and dynamic constraints.

Abstract Summary

Task and motion planning (TAMP) problems remain difficult even with full observability and object-centric states because discrete decisions are tightly coupled to geometric, kinematic, and dynamic constraints. Generalized TAMP addresses this difficulty by exploiting regularities across problem instances to reduce planning effort on new instances. However, existing methods require substantial TAMP-specific engineering. We investigate whether coding agents can automate this process by synthesizing programs that generalize across instances. Given a task description and simulator access, each agent chooses how to interact with the environment while developing a program within a fixed synthesis budget. The program is then frozen and evaluated on unseen instances. We evaluate Claude Code (Opus 5) and Codex (GPT-5.6 Sol and GPT-6 Astra) on 28 simulated environments from KinDER and PDDLStream, with object counts beyond those evaluated in the original benchmark. Across all program synthesis methods, we evaluate 980 generated programs on 100 held-out instances each, 98,000 evaluation episodes in total. Overall, we find that coding agents are surprisingly effective at generalized TAMP: all three agent configurations outperform hand-engineered planners, one-shot generation, and an LLM-based generalized planning baseline in mean success (56% to 95% versus 47% for the planners, on the 16 environments where a planner is available). As object counts grow, the agents' programs maintain higher success than the planner, using an order of magnitude less computation per instance on average. Logs show agents using interaction to calibrate physical models, test edge cases, and refine strategies. We release all code, including the full prompts given to the agents. These findings suggest that coding agents are a strong baseline for generalized TAMP.

Key Points

  • Task and motion planning (TAMP) problems remain difficult even with full observability and object-centric states because discrete decisions are tightly coupled to geometric, kinematic, and dynamic constraints.
  • Generalized TAMP addresses this difficulty by exploiting regularities across problem instances to reduce planning effort on new instances.
  • However, existing methods require substantial TAMP-specific engineering
  • We investigate whether coding agents can automate this process by synthesizing programs that generalize across instances.
  • Given a task description and simulator access, each agent chooses how to interact with the environment while developing a program within a fixed synthesis budget.

Problem Setting

We study finite-horizon, goal-directed Markov decision processes (MDPs). An MDP is a tuple $\mathcal{M}=\langle\mathcal{S},\mathcal{A},P,R,\rho,H\rangle$, where $\mathcal{S}$ and $\mathcal{A}$ are the state and action spaces, $P$ is the transition model, $R$ is a sparse reward function indicating goal achievement, $\rho$ is the initial-state distribution, and $H$ is the horizon.

Problem Setting — Coding Agents for Generalized Task and Motion Planning Problems
Fig. 3 : Left: custom tests reveal edge cases in StickButton. Opus writes a script that repeatedly calls reset with different seeds to find wall-adjacent sticks and high buttons for testing its program. The wall prevents a bottom grasp. A side grasp lets the robot move the stick away from the wall but leaves the high button out of reach; releasing and re-grasping from below provides the required height. Right: learning robot geometry through interaction. In Shelf, Opus probes the arm while holding a cube, fits its kinematic model to observed cube positions, and uses the calibrated model for inverse kinematics (IK). The initial model discrepancy is exaggerated for visibility.

Generalized TAMP with Coding Agents

We instantiate the synthesis method described in Section II using off-the-shelf coding agents from two providers: Claude Code running Opus 5 (high), and Codex running GPT-5.6 Sol (medium) and GPT-6 Astra (high). We choose Opus 5 as the model for all baselines, so that methods are compared with the same model.

Coding agents integrate frontier LLMs with harnesses that enable them to read files, write programs, and execute arbitrary commands.

Generalized TAMP with Coding Agents — Coding Agents for Generalized Task and Motion Planning Problems
Fig. 4 : Refining manipulation strategies during synthesis. In one Sol Blocked run, a revision adding a spare-block fallback changes success from 15% (A) to 56% (B); the pictured execution retrieves an alternative goal block from a distant table. The curve shows the held-out success rate of each commit, in commit order.

Experiments

We design experiments to answer the following questions about the efficacy and efficiency of AgenticGenPlan: Can agents write generalized programs for TAMP? What strategies do agents discover? What advantage does being “agentic” give? How efficient are the synthesized programs? Does access to environment source code help? We select the KinDER benchmark [ 20 ] as our primary benchmark.

This covers 25 environments, grouped into four families: Kinematic2D , Dynamic2D , Kinematic3D , and Dynamic3D .

Sources and Demonstrations

Figures are reproduced from Merler et al. See the full paper for experimental details and the project page, when the paper links one, for demonstrations and videos.

Share

Related Papers