Bimanual VLA: Reason, Act & Recover

Bimanual VLA: Reason. Act. Verify. Recover. is an end-to-end Physical AI system that transforms a natural-language dinner-table instruction into coordinated actions for two simulated SO-101 robotic arms in MuJoCo. The system uses Gemini as the cognitive planner to decompose a high-level instruction into structured manipulation tasks, dependencies, and workspace-aware left/right arm assignments. A dependency-aware bimanual scheduler then determines which tasks can be batched in parallel and which must execute sequentially. The selected actions are executed through numerical DLS inverse kinematics and reusable manipulation primitives including approach, grasp, lift, placement, and release. The system operates as a closed loop rather than a one-shot motion script. After manipulation, visual verification checks the resulting simulation state. When an action fails, the Recovery Engine diagnoses the failure, selects a recovery strategy, retries the action, and verifies the result again. Intel OpenVINO provides CPU-accelerated visual perception for edge-oriented inference. The system also includes a live observability dashboard showing the MuJoCo camera, cognitive task graph, bimanual scheduler, recovery telemetry, and hardware performance measurements. In the final end-to-end test, a real Gemini-generated six-task dinner-table plan was executed in MuJoCo with 6/6 tasks successfully completed. Two manipulation failures were detected and automatically recovered on the second attempt. On the target Intel Core i3-N305 system, the latest controlled benchmark measured 28.14 ms OpenVINO model-only inference and 37.72 ms for the complete perception pipeline, compared with 147.59 ms for the PyTorch CPU baseline, corresponding to a 3.91× speedup for the complete pipeline. The result is a reproducible closed-loop Physical AI architecture connecting language reasoning, bimanual planning, robotic control, Intel-accelerated perception, verification, and autonomous recovery.