Skip to content

FALLGUY

FALLGUY is a stunt-rehearsal simulator.

Core trick

Argon rebuilds a location video as a MuJoCo scene; a Unitree H1 humanoid performs the stunt while head-impact and joint-torque metrics are logged.

Tags

Description

FALLGUY is a stunt-rehearsal simulator. You upload a video or photo of a real location, and Gemini 4 Argon rebuilds it as a physics-ready, game-style 3D world, then checks it against the footage and refines it. You describe a stunt in plain English. Argon choreographs it, and a Unitree H1 humanoid performs it in MuJoCo while we trace the biomechanics: head G, HIC15, joint torques and impact forces. Argon then acts as the Assistant Stunt Director. It watches each take, reads the numbers and proposes a safer, better next take, so directors can test choreography changes in minutes with nobody at risk.
sudhersankv/ragdoll★ 00 forksPythonCreated Oct 25, 2026 · last push Oct 25, 2026

README excerpt

# FALLGUY: rehearse the stunt before anyone gets hurt

**Gemini 4 Argon × Google DeepMind hackathon, San Francisco**

Stunt choreography is one of the hardest jobs in film and live events. Teams spend weeks planning a single difficult stunt, but they rarely rehearse in the real location, and most of the shoot-day props aren't there during rehearsals. Performers end up improvising on the spot, and that's where injuries happen. AI video tools don't help: their clips look dramatic, but the physics is invented.

FALLGUY turns a phone video or photo of a real location into a physics-ready 3D world. A director describes a stunt in plain English. A Unitree H1 humanoid performs it in MuJoCo while we trace the biomechanics: head G, HIC15, joint torques, impact forces and spinal-load proxies. **Gemini 4 Argon** works as the Assistant Stunt Director throughout: it builds and checks the world, writes the choreography, watches every take and proposes a safer, better one.

Everything is viewed in the browser through a WebGL dashboard.

---

## The pipeline

```mermaid
flowchart LR
    A[Video or image<br/>of the location] --> B[Measure<br/>depth, floor, camera pan]
    B --> C[Argon maps the place<br/>objects, sizes, materials, camera]
    C --> D[Game-style 3D replica<br/>+ MuJoCo collision geometry]
    D --> E{Argon checks renders<br/>against the footage}
    E -- fix, up to 2 rounds --> D
    E --> F[Director approves<br/>or edits by prompt]
    F --> G[Director describes the stunt<br/

Team