Ghost Fighter

Ghost Fighter is a Physical AI / robot-sports platform that turns human movement into structured robot skills.

The demo path is:

human motion → UFB Studio / SONIC export → scored move card → 3D G1 replay → robot-sports arena

We score whether a human-created move is clean, expressive, and robot-executable enough to become a reusable robot-athlete skill.

cd web
npm install
npm run dev

Open:

• App: http://localhost:3000
• Move card / Skill Lab: http://localhost:3000/moves/ghost_jab_combo_sonic
• 3D Arena: http://localhost:3000/arena

The live Judge panel calls a local bridge, which sends move stats to the Judge uAgent and asks the Coach uAgent for fixes when needed. Before using the live Judge panel, run these in separate terminals:

cd agents
conda activate ghost-fighter-agents
python coach_agent.py

cd agents
conda activate ghost-fighter-agents
python judge_agent.py

cd agents
conda activate ghost-fighter-agents
uvicorn web_bridge:app --port 8010 --reload

Then start the web app with the bridge URL:

cd web
JUDGE_BRIDGE_URL=http://localhost:8010 npm run dev

The agents read agents/.env. Make sure it includes COACH_ADDRESS, JUDGE_ADDRESS, COACH_ENDPOINT, JUDGE_ENDPOINT, and BRIDGE_MODE=agent.

The arena can call fights with Deepgram TTS. Copy the example env file and add your key:

cp .env.local.example .env.local
# edit web/.env.local and set DEEPGRAM_API_KEY=...

Without a key, the app builds and runs normally — you just see “Announcer off” in the arena.

cd web
npm install
npm run build
npm start

If you see Can't resolve '@deepgram/sdk', run npm install in web/ first. The dependency is already listed in package.json.

1. Open /moves/ghost_jab_combo_sonic.
2. Show the move card: speed, power, smoothness, balance risk, recovery, deployability, and coach feedback (hover the i icons for definitions).
3. Show the 3D SONIC replay: G1 URDF driven by remapped joint_pos.csv trajectories.
4. Open /arena.
5. Use each player’s move buttons to trigger a 3D robot duel with HP bars, balance bars, hit effects, and knockback.
6. Toggle the announcer if DEEPGRAM_API_KEY is configured.
7. Mention the Unitree G1 assets loaded from web/public/models/g1_description and web/public/models/unitree_g1.

• / — dashboard and move library
• /ingest — upload a SONIC .zip or source video
• /moves/[id] — skill card, 3D replay, verification ladder
• /fighters/build — create a fighter loadout from move cards
• /arena — 3D robot-sports duel with health bars and move playback
• /leaderboard — top moves and fighters by deployability

The arena ships with a deterministic, seeded AI opponent. It satisfies the same per-frame controller contract as the human input handler, reuses the existing Move Card stats and arena loop/physics, and runs no LLM in the frame loop.

It thinks at three internal rates (mapped onto the arena tick by an adapter):

• Strategist (~3 Hz) — game-plan + online player-modeling (move-frequency, anti-spam counter-bias, intent: pressure / zone / counter / reset).
• Tactician (~15 Hz) — Utility-AI scoring of every Move Card now from the existing stats (range fit, whiff-punish/interrupt timing, safety, intent match, payoff − stamina − balance_risk − recovery) with seeded noise.
• Executor (per frame) — footwork toward preferred range, else commit the move via the arena's move-commit API.

Difficulty knobs (reaction_delay, optimal_prob, mistake_rate, adaptation, noise) are orthogonal to style (intent). Code lives in web/lib/enemy/.

Rusher, Zoner, Counter-Puncher, Adapter-Boss (web/lib/enemy/personas.ts).

Select one in the UI: open /arena and use the Player 2 AI dropdown. Choose a persona to let it pilot Player 2; "Manual (human)" returns control.

Headless: POST /api/arena/fight with { fighter_a, fighter_b, persona_b, seed } — when persona_b is set, that persona drives fighter_b. seed makes the match replayable.

Run a fighter's deck against the full persona pool through the real arena and get a win-rate + why-it-loses profile (and a readiness score blended with the existing Deployability):

curl -X POST http://localhost:3000/api/arena/evaluate \
  -H 'Content-Type: application/json' \
  -d '{ "fighter": "<fighter_id>", "matches": 5, "seed": 1 }'

cd web
npm test          # vitest, fully offline (seeded unit + real-arena integration)

SONIC zip exports store joints in IsaacLab/internal order (joint_0…joint_28). The web replay and trajectory API remap that to MuJoCo / Unitree SDK order before driving the G1 URDF (web/lib/g1Motion.ts).

Lafan CSV exports use a different layout (XYZ + QX QY QZ QW + 29 joints at 30fps). Use those for training pipelines like mjlab; the in-app replay expects the SONIC zip CSVs.

python3 scripts/analyze_motion.py

This reads the extracted SONIC CSVs and writes:

move_cards/ghost_jab_combo.json

• assets/motions/ghost_jab_combo_sonic.zip — Studio SONIC export for G1 deploy
• assets/motions/ghost_jab_combo_extracted/ — extracted SONIC CSVs
• assets/motions/block_sonic.zip — defensive block SONIC export
• assets/motions/block_extracted/ — extracted block motion CSVs
• web/public/models/unitree_g1/ — MuJoCo Menagerie Unitree G1 reference assets
• web/public/models/g1_description/ — Unitree ROS G1 URDF + meshes for browser rendering
• scripts/verify_with_gear_sonic.sh — prepares the motion folder for official GEAR-SONIC/MuJoCo verification

• Next.js 15 + React 19
• Three.js + URDFLoader for 3D replay and arena
• TypeScript scoring engine ported from scripts/analyze_motion.py
• JSON file store in web/data/ during local dev
• Optional Deepgram TTS for arena announcer (/api/tts)
• SONIC / G1 motion assets from UFB Studio

Ghost Fighter turns human moves into ranked, coachable, deployable robot skills. Studio retargets the motion; Ghost Fighter scores it, visualizes it on a G1 model, turns it into a move card, and lets it fight in a robot-sports arena.