AI

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

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 07/10/2026, 03:18:25 UTC.

Briefings

EmbodiedSmith is a framework for scalable embodied data generation using recursive self-improvement in simulation. It unifies asset, scene, and task generation to support autonomous creation and language-driven customization. The core mechanism is an agentic refinement loop where scene generation anticipates downstream tasks while task generation guides targeted scene edits, improving long-horizon task success. The framework supports diverse physical agents including mobile manipulators, humanoids, and dexterous hands, as well as deformable objects and fluids. Extensive experiments confirm that increased data diversity significantly improves generalization in downstream policy experiments.

Key points

• EmbodiedSmith unifies asset, scene, and task generation in a pipeline supporting autonomous creation and language-driven customization.
• The agentic refinement loop iteratively improves task generation success for long-horizon tasks by having scene generation anticipate downstream requirements.

Context

Scaling robotic foundation models requires diverse training data and reliable evaluation environments. Simulation offers a scalable solution, yet existing pipelines are constrained by predefined assets and skills, creating a disconnect between scene generation and task generation. EmbodiedSmith addresses this by unifying these components to support complex embodiments and physics.

Limitations

The framework supports mobile manipulators, humanoids, and dexterous hands, as well as deformable objects and fluids. Extensive experiments validate quality, diversity, and generation efficiency, but specific downstream policy results are limited to generalization improvements observed in experiments.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 24/10/2026, 23:43:40 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeig2a4wiguwa25tft4l5pdrv2ycbs7w3apd3mkmyu7vswkcimpgl7e