x-arm101

Two simulated SO-101 arms cooperate to set a dinner table in MuJoCo. One arm opens a drawer and retrieves cutlery, the arms perform a contact-based spoon hand-off, and a plate, cup, spoon and fork are placed on the table. Object poses, scene conditions and approach paths are randomized per seed, so each run requires the system to solve a slightly different task rather than replay a fixed trajectory. A cooperative scripted expert, built with inverse kinematics and contact-driven manipulation primitives, generates demonstrations for training. Across ten randomized evaluation seeds, the expert completes the full workflow with no monitored unintended collisions or numerical warnings. Demonstrations are exported as versioned LeRobot datasets and published as pinned Hugging Face releases for reproducibility. We fine-tuned SmolVLA on an Azure Tesla T4. The final multi-task dinner model was trained for up to 5,600 steps and early-stopped at the best validation checkpoint, step 4,749, with validation loss 0.0100. An earlier 20,000-step run on a 50-episode cup dataset established the training and evaluation pipeline. Policies are evaluated closed-loop on held-out seeds, with predicted actions validated before execution. Learned task success is not yet consistently reliable, and the scripted expert is used as the demonstration source and baseline rather than presented as the learned VLA. For Intel deployment, the fine-tuned policy exports to OpenVINO in FP16 and INT8 using static-shape prefix and denoise graphs. FP16 output differs from PyTorch by a maximum absolute action error of 0.00029, and the exported policy runs closed-loop in simulation. The pipeline supports CPU, iGPU and NPU benchmarking. Final Intel Core Ultra measurements are the remaining deployment step; current latency results are build-machine measurements only.