TABLEMATE · PRODUCT DEMO

From observation to motion.

An inspectable MuJoCo workstation for coordinating two SO-101 arms around a dinner table.

Watch the evidence ↓
Two simulated robot arms coordinating around dinner table objects

HOW IT WORKS

Bounded autonomy, visible decisions.

01 · Observe

Camera-grounded context

Overhead and wrist views feed a planner while joint observations keep the action interface explicit.

02 · Plan

Supervisor-owned steps

Supported skills, prerequisites, arm ownership, timeouts, and cancellation are checked before execution.

03 · Act

Contact-aware control

MuJoCo contacts represent grasps and hand-offs. Invalid or unsafe actions stop the run.

RECORDED EVIDENCE

A safe stop is a result.

These frames show a sealed contact-based teacher correction run. The repository preserves synchronized observations, actions, contacts, and outcome records for inspection.

Evidence boundary. The learned ACT candidate trained for 20,000 updates but stopped at the workbench-overlap guard during physical evaluation. Full learned dinner completion and Intel/OpenVINO execution are not claimed.

58-second MuJoCo teacher-evidence replay · source and manifests in the repository

REPRODUCE

Inspect the implementation.

Use the submission handoff for exact evidence boundaries, local commands, and artifact lineage. Explore the learning path to understand the system from first principles.