GS-Agent Puts a Physics Engine in the Multi-Agent Loop to Generate 4D Worlds From Text

arXiv 2607.21522·medium signal

A UMass Embodied AGI team built an end-to-end multi-agent framework that generates physically plausible, controllable 4D worlds from natural language by emulating the human workflow rather than learning generation end-to-end: the task decomposes into entity management (3D asset curation, material tuning, placement, motion control) and rendering configuration (camera and lighting), with specialized agents interacting with a physics engine via code and iterating on multimodal feedback. Results show rich interactions among liquids, deformable objects, and rigid bodies with cinematic camera and lighting control. The architectural lesson generalizes beyond graphics — agents writing code against a verifying simulator beats direct generation when physical plausibility is the constraint.