SentinelEdge: Bimanual VLA on Intel Core Ultra

SentinelEdge is a complete Physical AI system that tackles the three hardest bottlenecks in real-world robot deployment: multi-modal grounding, bimanual coordination, and edge inference latency. Two simulated SO-101 robotic arms operate inside a MuJoCo physics environment, executing complex multi-step dinner-table-setting tasks: opening drawers, retrieving objects, performing cross-arm hand-offs, and avoiding a dynamic sinusoidally oscillating obstacle. At the core is a Temporal Action Chunking Transformer (ACT) that fuses live camera pixels, real Sentence-Transformer language embeddings, and joint-action history through a 4-layer Cross-Modal Transformer, outputting independent 8-DOF trajectories for each arm. The policy is exported to Intel OpenVINO INT8 via NNCF Post-Training Quantization and benchmarked on Intel Core Ultra NPU at 85.34 FPS — 2.8× the headroom needed for 30Hz real-time control. Robustness is demonstrated across 10 randomized seeds with 6-axis domain randomization (placement, mass, friction, lighting, shape scale, floor texture). A Gradio dashboard with Speechmatics Batch ASR lets users speak commands that the robot executes and verbally confirms via Text-to-Speech. The project includes a multi-stage FSM reward system, 39-test pytest suite, ablation study notebook, HuggingFace Model Card, EMA kinematic action smoother for Sim2Real transfer, Foxglove Studio telemetry dashboard, and a full ROS 2 deployment node.