Dinner Claw

Dinner Claw is a bimanual manipulation stack for two simulated SO-101 arms in MuJoCo, built around one rule: a skill only succeeds if independent physics says so. Nothing scores for finishing its script. Each skill must show force-bearing contact from named gripper pads, a sustained lift with nothing outside the robot bearing the weight, and a stable released placement at resting height. Every skill ships with an open-gripper control that must score zero, and all four do, each rejected at its first physical gate. Measured on held-out seeds: drawer open 10/10, drawer-then-fork 10/10 chained in a single episode with no reset, no weld and no teleport, scripted cup place 10/10, camera-guided cup place 50/50, and a two-arm plate lift 9/10. The two-arm lift is the centrepiece. A single arm gripping a plate rim holds it 94 mm from its centre of mass, so it applies almost pure torque: it levers the plate onto its far rim and on six of ten seeds never gets it off the table. Two arms on opposing rims cancel that torque. The gate proves it with a 15-degree levelness bound that a pivot cannot satisfy however high it raises the plate's centre. The failing single-arm teacher is kept in the repository at 0/10 as the evidence for why the cell needs two arms. An ACT policy is trained with LeRobot and exported to OpenVINO FP32; an FP16 candidate failed numerical parity and was rejected rather than shipped. The project is deliberately explicit about what it has not done. There is no Intel Core Ultra result: all numbers here are Apple silicon and are diagnostics, not challenge results. No VLM drives the robot; the planner is a deterministic grammar that fails closed on anything unmeasured. The spoon handoff and pouring are unbuilt. Retracted results stay in the git history with the correction.