Cosmos3-Nano G1 + SONIC Forward Dynamics (vla480)
Post-training of the NVIDIA cosmos-framework Cosmos3-Nano model on Unitree G1 + GEAR-SONIC data (Humanoid-Everyday) as a forward dynamics model. Given a first ego frame, 78-D action tokens, and a task prompt, it renders future ego video. Resolution 640x480, 3 epochs.
Results
In-distribution rendering verified via autoregressive chunk chaining (full episode reproduced). Out-of-distribution motions such as kneeling and novel scenes fail due to limited training data, since a renderer cannot produce motions absent from the training set. iter_000004911 is the final checkpoint (3 epochs) and includes model, optim, scheduler, and trainer states, so it is resumable.
License and Attribution
Base model: NVIDIA Cosmos-Predict2.5 (Cosmos3-Nano). OpenMDW-1.1. Commercial use allowed with attribution.
