OpenDine

OpenDine delivers an end to end, closed loop Physical AI system for autonomous bimanual table preparation inside the MuJoCo 3 physics simulation engine. The platform coordinates dual 6-DoF SO-101 robotic manipulators to execute a complete 18 stage physical manipulation workflow conditioned on natural language instructions and multi-camera inputs (an overhead workspace view and dual in-hand wrist streams). Key coordination milestones include center plate positioning, organizer cutlery retrieval with coordinated in-air center hand-offs, stabilizing a beverage bottle to unscrew its cap 90° counter-clockwise, and streaming fluid directly into a ceramic mug before retracting to standby. The underlying SmolVLA policy combines a SigLIP vision encoder, SmolLM2-500M language backbone, and a flow-matching action head. Optimized for edge robotics with Intel OpenVINO, the pipeline utilizes an NNCF INT8-quantized vision tower compressed 3.9× down to 25.8 MB with minimal action drift (0.0310 rad). Benchmarked inside an Intel ready Podman container, the CPU runtime achieves 289.34 FPS with a 3.03 ms mean latency, delivering over 28× headroom above the 10 Hz simulation loop. Evaluated across 10 randomized seeds with mass, position, and surface friction perturbations, OpenDine maintained a 100% task success rate. The project is delivered alongside an interactive browser-based app featuring native 3D drag-to-orbit camera controls, dynamic precision toggling, and live hardware telemetry.