Coding agents generate robot training data that transfers to a real robot

single source· 1 articles · confidence: low · first seen 2026-09-22 20:00 UTC

What this means for you

Nothing to act on today: this is a preprint, and the abstract names no code release, no model weights and no evaluation numbers. Worth tracking if you collect robot demonstrations by hand — the claim is that verified simulation solutions can substitute for them, at the cost of heavy per-task iteration.

Researchers posted a preprint describing EMBODIEDSWE, a framework that turns coding agents — systems that write and run their own code — loose on robotics tasks in simulation, then converts their solutions into training data for robot policies. The benchmark, EMBODIEDSWE-BENCH, spans contact-rich manipulation, deformable objects and tasks requiring up to thirty minutes of continuous interaction. Frontier agents solved many of them and reused solutions across tasks and robot bodies, but needed heavy iteration and produced answers specific to individual instances. A vision-language-action model (mapping images and instructions to motor commands) fine-tuned only on agent-generated simulation data completed a long-horizon task on a physical robot.

Key facts

  • ·EMBODIEDSWE-BENCH is a simulation benchmark for coding agents covering contact-rich manipulation, deformable objects, and tasks requiring up to thirty minutes of continuous interaction; the preprint is arXiv 2609.27308, posted 22 September 2026. source
  • ·The paper reports that frontier coding agents solved long-horizon benchmark tasks and transferred prior solutions across both tasks and robot embodiments. source
  • ·Agent solutions required substantial iterative interaction and were typically specialised to individual task instances, according to the paper. source
  • ·EMBODIEDSWE-GEN expands a single coding-agent solution into large diverse trajectories for training a vision-language-action model. source
  • ·The paper states VLA performance improved with more generated demonstrations, and agent-aided diversification improved generalisation to held-out task variations. source
  • ·A VLA fine-tuned solely on coding-agent-generated simulation demonstrations completed a long-horizon task on a real robot. source

What the sources say

Sources

The original reporting. Follow these — they did the work.

← the wire