Headset
On the headset, open this website, then menu, TRAINER ON YOUR COMPUTER, and pick this code.
Inbox
Scenario packs from the headset: falls and stuck moments, captured with the room, terrain and the commands that led there. Tick the ones to train on.
Motion sets
Scene
Runs the pack in Isaac Sim on this PC: Pip under PhysX, beside an orange ghost replaying MuJoCo's run of the same pack. Watch it in the stream.
Train
Local fine-tune: evolution strategies on the walking policy's last layer, in MuJoCo on this PC, with shoves and noise. Checked on held-out seeds and on plain floor before it is released.
Every new brain is then cross-checked against the shipped one in MuJoCo, MuJoCo with the randomisation, and Newton.
Cross-check
Brains on the ticked packs in every physics engine set up on this PC: MuJoCo as on the headset, MuJoCo with the randomisation (several seeds), Newton's MuJoCo-Warp on the GPU (MuJoCo's contacts, and Newton's own as NVIDIA trains with), and Isaac Sim's PhysX. Falls and distance per brain.
Releases
Export
Take the ticked test cases (and your motion sets) to any simulator or trainer: MuJoCo scenes of your room, meshes for other engines, the recorded run as a dataset, the policies and a Python environment, in one .zip. Brains trained elsewhere come back as a release folder.