EmbodiedSWE: Coding Agents for Long Horizon Dexterous Robotics
Haoxiang You, Zeyu Shen, Yilang Liu, Zhicheng Zheng, Lihan Zha, Kashu Yamazaki, Mingtong Zhang, Suning Huang and 11 others
ByteDance Seed · Yale University · Princeton University · Carnegie Mellon University · Stanford University · University of California, Los Angeles
Published September 23, 2026 · on the feed September 24, 2026 · ▲ 3 · arXiv 2609.27308
The abstract
We study coding agents for long-horizon, dexterous robotics and ask whether their solutions can provide scalable supervision for learning general robot policies. To test this, we develop EMBODIEDSWE-BENCH, a simulation benchmark for coding agents spanning contact-rich manipulation, deformable objects, and long-horizon tasks requiring up to half an hour of continuous interaction. We find that frontier coding agents can solve complex long-horizon tasks and transfer prior solutions across both tasks and embodiments. We also design supporting tools that help agents more effectively solve these tasks. However, the resulting solutions require substantial iterative interaction and are typically specialized to individual task instances. We therefore introduce EMBODIEDSWE-GEN, which expands a single solution from coding agent into large diverse trajectories for training a VLA. VLA performance improves with more generated demonstrations, and agent-aided diversification improves generalization to held-out task variations. We also show that a VLA finetuned solely on coding-agent-generated simulation demonstrations completes a long-horizon task on real robot. Together, our framework uses coding agents to solve complex robotics tasks and turn verified solutions into scalable supervision for robot policies.
The Brief
2:25A short podcast summary of the paper. Written and voiced by AI.
Transcript
EmbodiedSWE. Coding Agents for Long Horizon Dexterous Robotics. By Haoxiang You and eighteen others, from ByteDance Seed, Yale University, and five other institutions. Published on the twenty-third of September. Its code repository is worked out: about three hundred source files, with a readme and a licence.
The question is whether coding agents can handle long-horizon, dexterous robotics, and whether their solutions can then provide scalable supervision for learning general robot policies. Testing that needs somewhere hard enough to try. The idea being probed is that writing code could stand in for demonstrations. If an agent can program a robot through a task, perhaps what it produces can be used to teach a learned policy.
EMBODIEDSWE-BENCH is a simulation benchmark for coding agents. It spans contact- rich manipulation, deformable objects, and long-horizon tasks that require up to half an hour of continuous interaction. On it, frontier coding agents can solve complex long-horizon tasks, and they transfer prior solutions across tasks and across embodiments. The authors also design supporting tools that help agents solve these tasks more effectively. The benchmark therefore answers the first half of the question. Coding agents can get there, even on tasks this long and this physical.
There is a catch. The resulting solutions take substantial iterative interaction to produce, and they are typically specialized to individual task instances. So the authors introduce EMBODIEDSWE-GEN, which takes a single solution from a coding agent and expands it into a large, diverse set of trajectories for training a VLA. One expensive, narrow success becomes many varied examples. That is the step that turns a solution into supervision, since a learned policy needs breadth that a single solved instance does not have.
VLA performance improves as more demonstrations are generated, and diversification aided by the agent improves generalization to held-out variations of a task. A VLA fine-tuned solely on simulation demonstrations generated by coding agents completes a long-horizon task on a real robot. Those results trace the whole path the authors set out. Agents solve in simulation, their solutions are multiplied, a policy learns from the copies, and the policy runs on hardware.
The framework turns solving into teaching. A coding agent's verified solution is costly and narrow on its own. Expanded into many varied demonstrations, it becomes supervision a robot policy can learn from, and that expansion is where the scalability comes from.
Paper Quest · 7 chapters · about 39 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Robots need demonstrations, and demonstrations are expensive. That is the wall.
You want a general robot policy that can do long, dexterous jobs, like tasks needing half an hour of continuous contact-rich work. Such policies learn from demonstrations.
The road ahead
- 1 · The problem ~6 min Robot policies learn long, dexterous jobs from demonstrations, but good demonstrations of such jobs are slow, costly and sometimes nearly impossible to collect.
- 2 · What came before ~3 min
- 3 · The key idea ~8 min Coding agents write robot programs that solve tasks in simulation; one costly, narrow solution is then expanded into many varied trajectories that train a VLA.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min An agent solves each task in simulation, one verified solution is expanded into diverse trajectories, a VLA is fine-tuned on them, and it runs on a real robot.
- 6 · The experiment ~6 min You’ll guess: Fine-tuned only on coding-agent-generated simulation demonstrations, could a VLA complete a long-horizon task on a real robot?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.