From English to robot simulation, automatically

Creating robot simulation scenes requires specialized skills in 3D modeling, physics specification, and spatial layout—work that slows down simulation-based learning pipelines. SR-Platform chains four components: an LLM orchestrator that plans scenes, an asset forge that retrieves or generates 3D objects via CadQuery synthesis, a layout architect that assigns poses and enforces constraints, and a bridge that produces the final MJCF file. The system runs as a nine-service Docker stack and has processed 611 production jobs over 30 days, generating five-object scenes in 50 seconds median latency (30–40 seconds with cached assets). The asset forge recovers from failures with 11.3% retry rate. Code and deployment architecture are production-ready.