FORGE

Inspiration

A robot can finish its move and still disturb the object beside it. A success counter misses the incident, while a replay alone cannot tell an engineer which change would help. FORGE connects the failure, a bounded experiment and the next release decision.

What it does

FORGE is a reliability workbench for an original simulated Cartesian gantry. Real MuJoCo contact physics determines outcomes. A compact 650-parameter network receives eight instrumented measurements and predicts two waypoint targets for a fixed controller. The inspection camera helps a person understand the incident; the policy does not consume camera pixels.

The demo opens a retained failure and a completed diagnostic. NVIDIA Nemotron Ultra on Nebius proposed reducing the protected object's height from 170 mm to 65 mm. Both branches used the same frozen policy and paired random variations. In 20 fresh paired trials, successes changed from 0/20 to 20/20: 20 rescues and no observed regressions in that sample. The source-seed replay is separate from those fresh trials.

Height also changes the policy's measurements. This is a scene diagnostic, not a claim that the experiment trained or improved the policy.

How I built it

React, TypeScript and Three.js present synchronized replays above Python/FastAPI, SQLite and immutable evidence. Owned workers run actual physics, collect corrections and train compact policy weights. Exact source, checkpoint, proposal and outcome identities travel with executable exports.

Nebius Token Factory served Nemotron Ultra for the demonstrated intervention proposal. Fireworks Ultra and Together Nemotron Super VL supported earlier investigation. The physics oracle, not a model explanation, decides outcomes.

Challenges and what I learned

The most useful result was also inconvenient. In the separate registered release study, the minimized nominee passed 297/300 shifted cases. Both correction comparators passed 300/300. The required advantage was not met, so FORGE keeps the decision: do not promote.

Separating that study from the successful scene diagnostic became a core product requirement. Engineers need to know which question a result answers. Failed runs and rejected proposals therefore remain in the record.

What comes next

The film shows actual application modules and recorded simulation replay for more than one minute. No physical robot, hardware transfer, GPU execution, GR00T rollout or independent engineer validation is claimed. A pinned GR00T/LIBERO worker package is included with its pending gates visible. Next comes qualifying that path and testing the evidence workflow with larger policies and real hardware.

Built With

• fastapi
• mujoco
• nebius
• nvidia-nemotron
• python
• react
• sqlite
• three.js
• typescript