VoxHands

VoxHands is a voice-first dual-arm tabletop assistant that separates language from action. A Groq-powered tool loop turns a command such as "set the table for two" into a bounded set of workcell tools, never raw motor commands. A deterministic Python validator enforces workcell bounds, collision rules, and no-go safety zones before anything moves, refusing unsafe requests outright. The Python simulation owns authoritative object and arm state, while the browser renders the workcell and confirms contact through Rapier3D physics at 60 Hz. The deployed service runs a real MuJoCo backend: two SO-101-style six-DOF arms driven by Jacobian inverse kinematics, kinematic grasp and release, and gravity-settled placement, with overhead and operator camera views. MuJoCo's offscreen renderer needs an EGL/OSMesa GL stack that Render's native Python runtime cannot install, so the blueprint builds a Docker image (Python 3.13, mujoco, Pillow, OpenVINO); the robot adapter now reports available live instead of falling back to the deterministic simulator. VoxHands also carries a real Intel optimization path, not a stand-in: the policy network is exported to a genuine OpenVINO IR, quantized to INT8 with NNCF post-training quantization, and benchmarked on CPU. INT8 reaches 1.21x to 1.71x speedup on compute-bound models and 5,003 inferences/second asynchronous throughput, while we state plainly that the tiny workcell policy is dispatch-bound and gains nothing from INT8. All figures are measured on the live host. That honesty is the point: we report what is verified and gate what is not, with GPU and NPU acceleration explicitly unverified until it runs on Intel Core Ultra hardware. The result is a reproducible, MIT-licensed testbed for physical AI where optional backends, MuJoCo, OpenVINO vision, and browser physics, activate when present and degrade gracefully when absent. Grasping and placement remain simplified and no physical robot is driven: everything executes in simulation.