Research map / Multi-agent systems and games / LLM agents

SimWorld Studio: Automatic Environment Generation with Evolving Coding Agent for Embodied Agent Learning

NeurIPS 2026 · Multi-agent systems and games › LLM agents · approach: LLM multi-agent collaboration and safety · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces SimWorld Studio, a UE5-based platform where a self-evolving tool/skill-augmented coding agent (SimCoder) generates verified 3D embodied environments exported as Gym envs, and co-evolves difficulty with an embodied navigation agent.

Problem

Embodied agents lack abundant, diverse, automatically generated interactive 3D environments with verifiable tasks and standard learning interfaces; existing LLM scene generators produce static scenes and simulators rely on manual or template-based scenes.

Approach

A coding agent writes engine-level UE5 code under MCP tools, uses rule and VLM verifier feedback to revise scenes, authors new tools and skills into its library (self-evolution), exports Gymnasium navigation tasks, and adapts environment difficulty from embodied agent success (co-evolution).

Key results

• Co-evolution reaches 90% test success rate on SimWorld-MMNav, an 18-point gain over fixed-environment training (72%) and a 40-point gain over the untrained agent (50%).
• Opus 4.6 achieves the highest scene-generation average (0.77 text-to-scene, 0.79 image+text, 0.75 editing), with collision-free rates at or above 0.98 for Opus and Sonnet.
• Adding self-evolution to the coding agent raises scene generation quality by +0.21 over static configurations (ablation starting from 0.16 vanilla).

Surprising findings

• Qwen3.5-9B baseline achieves 0.40% SR on PointNav and 0.00% on ObjectNav before memory, indicating near-zero zero-shot navigation for small models.

Limitations

• Effectiveness bounded by the underlying coding agent's spatial reasoning capability.
• Co-evolution demonstrated only in a single navigation case study with one embodied model (Qwen3.5-9B).

Experimental setup

Datasets: SimWorld-generated navigation episodes (1.2K per task)

Benchmarks: SimWorld Studio held-out navigation episodes (329 per task) SimWorld-MMNav

Metrics: Success Rate SPL SoftSPL nDTW rule-based validity metrics VLM-judged fidelity and aesthetics

Baselines: Claude Sonnet 4.6 Qwen3.5-27B Qwen3.5-9B Qwen3.5-2B fixed-difficulty environment random-difficulty environment untrained base model

Builds on: SimWorld Model Context Protocol PPO ReAct ExpeL Generative Agents GEPA EnvGen

Closest papers by meaning

• VirtualEnv: A Platform for Embodied AI Research (AAAI 2026 · similarity 0.89)
• Learning Interactive Real-World Simulators (ICLR 2024 · similarity 0.89)
• Exploration-Driven Generative Interactive Environments (CVPR 2025 · similarity 0.89)
• Learning 3D Persistent Embodied World Models (NeurIPS 2025 · similarity 0.88)
• Embodied Multi-Modal Agent trained by an LLM from a Parallel TextWorld (CVPR 2024 · similarity 0.87)
• LabUtopia: High-Fidelity Simulation and Hierarchical Benchmark for Scientific Embodied Agents (NeurIPS 2025 · similarity 0.87)