EmbodiedSWE: Coding Agents for Long Horizon Dexterous Robotics

This paper investigates coding agents for long-horizon, dexterous robotics using EMBODIEDSWE-BENCH, a simulation benchmark requiring up to half an hour of continuous interaction. The authors find that frontier coding agents solve complex tasks and transfer solutions across embodiments. To leverage this, they introduce EMBODIEDSWE-GEN, which translates agent solutions into diverse trajectories to train Vision-Language-Action (VLA) models. Experiments demonstrate that a VLA finetuned exclusively on these simulated demonstrations successfully executes a long-horizon task on a real robot.

Frontier coding agents can solve complex long-horizon tasks requiring up to half an hour of continuous interaction.

EMBODIEDSWE-GEN expands a single solution from a coding agent into large diverse trajectories for training a VLA.

A VLA finetuned solely on coding-agent-generated simulation demonstrations completes a long-horizon task on a real robot.