TwinGuard: Closed-Loop Bimanual Physical AI

TwinGuard is an autonomous bimanual Physical AI system built for the Intel Physical AI Online Challenge ("Setting Up a Dinner Table"). Operating two 6-DOF SO-101 robotic arms in MuJoCo, TwinGuard solves the critical flaw of open-loop robotics: catastrophic task failure when unmodeled physical slips or unexpected contacts occur. TwinGuard introduces a closed-loop hierarchical autonomy architecture pairing multimodal vision-language reasoning with continuous physical and visual verification, autonomous self-healing recovery, and Intel OpenVINO edge acceleration. Key Features & Intel Hardware Placement: 1. Heterogeneous Intel Core Ultra Architecture: - Intel NPU (AI Boost): Runs OpenVINO INT8 (NNCF) quantized neural vision and Intel Anomalib-aligned visual anomaly embeddings with 0.58ms latency (~1,700 FPS). - Intel Arc iGPU: Drives parallel off-screen camera rendering and asynchronous frame processing via OpenVINO Async Engine (1,439 FPS). - Intel CPU (P/E-Cores): Executes 500 Hz MuJoCo physics, Damped Least Squares inverse kinematics, and contact force verification (5.3x real-time). 2. Coordinated Bimanual Manipulation & Handover: Implements 8 motion primitives including an explicit mid-air bimanual handover between arms and synchronized dual-arm pouring, achieving 0 collisions across all runs. Exposes the official Hugging Face LeRobot bi_so101_follower 12-DOF action space and SmolVLA observation schema. 3. Closed-Loop Self-Healing Recovery: Across 10 randomized seeds with controlled physical grasp slips injected into Seeds 2 and 7, TwinGuard’s SafetyVerifier and VisualGuard immediately detect anomalies, visually re-scan the workspace, dynamically replan, and achieve a 100% post-recovery task completion rate. 4. Production Engineering: Backed by 112 passing automated unit tests, clean configuration files without magic numbers, and an isolated reproducible Docker container.