RecoverVLA Dual-Arm Dining

RecoverVLA is an Intel Physical AI online-track system: two simulated SO-101 arms in MuJoCo that turn a natural-language table-setting instruction into a coordinated bimanual skill plan, execute it with retries, and report what succeeded and what failed. The team builds and runs the stack on an Intel Core laptop. A typed instruction becomes an explicit plan for both arms: open the drawer, grasp and place the plate, grasp the mug, grasp the bottle, pour, place the mug, and verify the final scene. Every skill names an arm, a target, a success check, and a retry budget. Seeded scenes randomize object pose, mass, friction, and lighting without making the task physically invalid. Results are machine-readable JSON (task success, subtask progress, recovery count, failures) so later OpenVINO latency, throughput, precision, and device fields can be filled on the same Intel machine instead of being invented for a demo. What has been verified in the MuJoCo dining scene: a closed-gripper paddle opened the drawer by contact in 10/10 seeds (100-109), and a two-jaw pinch lifted a central plate tab in 10/10 of the same seeds. Seed 100 then placed the plate, lifted the mug with the right arm, and lifted the bottle by a neck tab with the left. Pouring is not solved: beads leave the bottle and land on the table, not in the mug. The default evaluation command remains a seeded dry-run until the MuJoCo executor is wired through as the default path.