EmbodiedSmith: Scaling Embodied Data through Recursive Self-Improvement Flywheel in Simulation

EmbodiedSmith. Scaling Embodied Data through Recursive Self-Improvement Flywheel in Simulation. By Yikai Qin and fifteen others, from the Hong Kong University of Science and Technology, Guangzhou, Xiaomi Robotics, and eight other institutions. Published on the sixth of October. No code repository is linked to it.

Robotic foundation models need diverse training data to scale, and they need reliable environments to be evaluated in. Simulation offers both at scale. The pipelines that generate simulated data, however, are held back in three ways. They depend on predefined assets and skills. Scene generation and task generation are disconnected from each other. And support for complex embodiments and complex physics is limited. Each of these narrows what the data can contain. Fixed assets and skills cap the variety. A scene built without regard to the task it will host may not suit that task. And a pipeline limited in bodies and physics can only produce data about the behaviours it can simulate.

EmbodiedSmith is a framework for generating embodied data at scale through recursive self-improvement. It unifies the generation of assets, scenes, and tasks in one pipeline, which can run autonomously or be customised through language. At its core is an agentic refinement loop in which scenes and tasks improve each other.

The loop runs in both directions. Scene generation anticipates what the downstream tasks will require. Task generation, in turn, guides targeted edits to the scene. Iterating between the two lets each one improve the other, and this joint refinement raises the success of task generation, including for long-horizon tasks. The framework also covers a broad range of bodies and physics: mobile manipulators, humanoids, and dexterous hands, along with interactions involving deformable objects and fluids. That widens the behaviours and physical phenomena the generated data can represent.

The evidence comes in two layers. Extensive experiments first validate the data itself: its quality, its diversity, and how efficiently it is generated. Downstream policy experiments then ask whether that diversity pays off for the robots trained on it, and show that greater data diversity improves generalisation. That second layer is the one the whole pipeline is built to serve, since the point of scaling data is the policies it produces.

Taken together, the result is a flexible simulation engine meant for both pretraining and evaluating robots. The connecting thread is the loop: rather than building a scene and then hoping a task fits it, EmbodiedSmith lets the task shape the scene and the scene anticipate the task, so the data improves as it is produced.