AssemblyWorld: Rethinking 3D Assembly with General-Purpose Agents
This preprint introduces AssemblyWorld, an interactive 3D simulation where general-purpose multimodal agents assemble rigid parts through rendered views and pose-editing tools. The strongest evaluated agent, GPT-6 Astra, reaches 80.9% part accuracy but only 59.4% complete-assembly success across 100 tasks spanning furniture, industrial parts, and broken-object reassembly. It matters because off-the-shelf agents can reason about assembly without task-specific fine-tuning, but precision and execution reliability still fall short of deployment needs.