SimWorld Studio: Automatic Environment Generation with Evolving Coding Agent for Embodied Agent Learning
Researchers have introduced SimWorld Studio, an open-source platform built on Unreal Engine 5 designed to address the scarcity of diverse, automatically generated 3D environments for embodied agent learning. Unlike existing simulators that rely on manual crafting or static scenes, SimWorld Studio features SimCoder, a tool-augmented coding agent that constructs physically grounded 3D worlds from language or image instructions. SimCoder self-evolves by utilizing verifier feedback, such as compilation errors and physics checks, to refine environments and expand its library of reusable skills. The platform exports these worlds as Gym-style environments, facilitating interactive training. A key innovation is the co-evolution mechanism, where agent performance feedback guides the generation of adaptive curricula tailored to the learner's capability frontier. Case studies in embodied navigation demonstrate that this approach significantly enhances generation reliability and agent performance. Specifically, co-evolution yielded an 18-point success rate improvement over fixed-environment learning and a 40-point gain compared to untrained agents, highlighting its potential to accelerate advancements in embodied AI through scalable, dynamic simulation environments.
Wire timeline
SimWorld Studio: Automatic Environment Generation with Evolving Coding Agent for Embodied Agent Learning
Researchers have introduced SimWorld Studio, an open-source platform built on Unreal Engine 5 designed to address the scarcity of diverse, automatically generated 3D environments for embodied agent learning. Unlike existing simulators that rely on manual crafting or static scenes, SimWorld Studio features SimCoder, a tool-augmented coding agent that constructs physically grounded 3D worlds from language or image instructions. SimCoder self-evolves by utilizing verifier feedback, such as compilation errors and physics checks, to refine environments and expand its library of reusable skills. The platform exports these worlds as Gym-style environments, facilitating interactive training. A key innovation is the co-evolution mechanism, where agent performance feedback guides the generation of adaptive curricula tailored to the learner's capability frontier. Case studies in embodied navigation demonstrate that this approach significantly enhances generation reliability and agent performance. Specifically, co-evolution yielded an 18-point success rate improvement over fixed-environment learning and a 40-point gain compared to untrained agents, highlighting its potential to accelerate advancements in embodied AI through scalable, dynamic simulation environments.
cs.AI updates on arXiv.org