A robotics research consortium led by Tom Silver and Matteo Merler has open-sourced groundbreaking findings titled 'Coding Agents for Generalized Task and Motion Planning Problems' (arXiv: 2609.30233, GitHub: tomsilver/robocode). Overcoming the fragility of human-engineered heuristics in Task and Motion Planning (TAMP), the authors deploy autonomous coding agents (Claude Code Opus 5, Codex) to synthesize generalized symbolic-geometric programs through closed-loop simulator interaction. Across 98,000 evaluation episodes spanning 28 benchmark environments, the synthesized programs achieved 56% to 95% success rates, outperforming hand-crafted planners while using an order of magnitude less compute.
Key Takeaways
- ✓Replaces handcrafted robotic heuristics by deploying autonomous coding agents to synthesize generalized TAMP programs.
- ✓Agents autonomously interact with physics simulators to stress-test edge cases, calibrate dynamics, and refine logic.
- ✓Evaluated across 28 simulation environments and 98,000 episodes, lifting mean task success from 47% to up to 95%.
- ✓Maintains superior robustness as object counts scale while requiring an order of magnitude less compute per instance.
- ✓Full benchmark harness, agent interaction logs, and prompts open-sourced on GitHub under Apache-compatible license.
Key Decision Metrics at a Glance
Heavy Claude Code use: compare subscription limits and API bills
Compare 40 dev plans & simulate token costs vs $20/mo subscriptions
Project Links & Resources
Direct AccessIn-Depth Technical Analysis
Core Background & Industry Pain Points
Task and Motion Planning (TAMP) is foundational for autonomous robotics, dictating how manipulators sequence discrete sub-goals while honoring continuous kinematic, collision, and dynamic constraints. Yet generalized TAMP has historically demanded months of handcrafted engineering. Existing symbolic planners suffer from combinatorial explosion when scaling to multi-object cluttered environments, causing catastrophic planning latency or deadlocks.
Architecture Highlights & Internals
The authors demonstrate that autonomous coding agents (Claude Code powered by Opus 5 and Codex) can automate TAMP engineering. Armed with task specifications and physics simulator access within a fixed compute budget, the agent writes, tests, and refines generalized Python programs. It utilizes environment feedback to calibrate physics models, stress-test collision tolerances, and resolve dynamic edge cases. Once finalized, the program is frozen and deployed zero-shot across unseen physical environments.
Authoritative Benchmarks & Measured Scores
Across 980 generated programs evaluated across 98,000 episodes on 28 KinDER and PDDLStream environments: synthesized programs achieved 56% to 95% task success, vastly eclipsing the 47% baseline of hand-crafted planners. Crucially, as obstacle and object counts scaled, the coding agents' programs preserved high success rates while consuming an order of magnitude less runtime compute per episode.
Developer Hands-on Guide
Robotics researchers can clone https://github.com/tomsilver/robocode to run the reproducible evaluation suite, inspect agent reasoning transcripts, and leverage the included prompt templates for custom simulators.
Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.
Discussion & Comments
0Sign in to join the discussion
Connect with AI developers to exchange benchmark insights.