Twenty million people play chess daily — but we've reduced humanity's richest game to clicking squares on a screen.
Your friend from across the country. The world's top engines.
Chess engines have been world-class for thirty years.
Your wooden board has none of them.
BetaGo closes that gap.
The reBot DevArm — Damiao motors over USB-CAN, a parallel gripper, and an overhead camera watching the board. That's the entire physical setup.
The arm is the same one rendered live in the UI. The URDF that drives the on-screen 3D twin is the URDF we use for inverse kinematics — one source of truth.
LIVE · BOARD CAM + ARM
Everything else orbits it. Eval bar on the left, Lichess-style. Multi-PV arrows on the squares — emerald for the engine's pick, lime for the runners-up. A Decision Trace strip below ranks the top five.
Toggle either side between human and AI, pick a model (here Claude Opus 4.7 on both), watch eval / depth / PV stream live. State badges — THINKING, HANDS-ON, YOUR TURN — make ownership unmistakable.
Each side has a floating console rendering the URDF arm, mirroring the physical robot's joints in real time.
A floating console docks the overhead camera (RAW / FEN / DIFF overlays) in live play, or run.mp4 with the stage timeline in Replay mode.
Quick chips — hint, explain that move, take it back, reset — plus a free-text prompt for the coach.
Each layer is independent and replaceable. The planner is a state reconciler between three sources of truth — the engine's bestmove, the camera's observed FEN, and the arm's actual execution.
Standard chess engine protocol — swap Stockfish for Lc0, Komodo, anything that speaks UCI.
Overhead camera frames go to a VLM that returns the observed FEN. Authoritative for human moves.
UCI move resolves to source/dest squares, IK plans the trajectory, the arm executes and verifies.
A Star Arm 102-LD leader drives the B601-DM follower while two cameras and seven joints log every frame. Each clip is one isolated pick-and-place — 50 episodes per piece type, split white / black — pushed straight to a private Hugging Face dataset for Pi0 fine-tuning.
The same URDF that drives the on-screen arm twin powers a full PyBullet simulator. The planner ships every trajectory through sim first — collision-checked, joint-limit-clamped — before a single Damiao motor turns.
Captured straight from the workshop: the reBot DevArm picking and placing a piece while the planner watches the board.
The planner walks the arm through ten phases in sequence — the same ladder rendered on the UI. Vision verifies the piece actually landed before the digital board commits.
Camera pulls a fresh frame; the arm holds rest pose.
UCI for the engine, REST + WebSocket between containers, URDF shared between sim and hardware. Every layer can be replaced without touching the others.
Joint stream, arm-stage events, vision-verified completions. The on-screen arm reflects what the physical arm actually did, not what we hoped it would do.
Engine bestmove, observed FEN, arm execution. The planner mediates. The LLM only enters when the three disagree.
Built in two days at the Physical AI Hackathon · May 9–10, 2026.