robium
robium / applications

Quadruped locomotion

Train a Unitree Go2 walking policy in Isaac Lab, compare real checkpoints, and optionally pilot it in a private browser session.

experimental Source ↗
reinforcement-learningisaac-labisaac-simgo2gpu-cloud
A simulated Unitree Go2 walking across a flat gray Isaac Lab scene
A simulated Unitree Go2 walking across a flat gray Isaac Lab scene

01 · Recorded interactive · local preview

Open interactive playback

Switch between saved policies and watch how the gait changes without starting a GPU.

  • No model training or application installation
Open checkpoint demo

02 · Remote Isaac Sim

Live Isaac Sim on RunPod

Open the real Isaac Sim viewport and controller from a compatible remote GPU session.

  • Live simulator frames
  • Velocity controls and checkpoint loading
  • Uses the recovered capability-scoped controller
Open managed simulator

03 · Compatible GPU host

Run it on a GPU

Clone the reference application on a compatible remote GPU host and run its preflight before training.

git clone https://github.com/robium-ai/robium-apps
cd robium-apps/quadruped-locomotion
make smoke
Open application Read guide

Recorded checkpoint explorer

Scrub from unstable motion to a walking gait

These clips are byte-matched to the recovered run archive. Select a checkpoint to see how the same velocity task changed during training.

training time
32 min
saved checkpoints
21
final mean reward
34.97

Historical baseline from 2026-07-27. A fresh current-stack evaluation remains required before publication.

iteration 1999