DreamDojo 2B G1 + SONIC (Continued Post-training)
Continued post-training of the NVIDIA DreamDojo 2B G1 world model on our own Unitree G1 + GEAR-SONIC data (box + Humanoid-Everyday, 258 episodes). Whole-body joint angles (wbc, 43-D) are placed in the G1 slice of the 384-D shared action space, inheriting the pretrained action embedder for few-shot adaptation.
Results
Adapts within about 1000 steps. Previously out-of-distribution box (kneeling) renders coherently. Short clips (13 frames) reach SSIM 0.72. Each iter folder contains model, optim, scheduler, and trainer states, so training can be resumed.
Checkpoints
- iter_000070000: latest, for resume.
- iter_000040000: mid-training, safe fallback teacher.
License and Attribution
Base model: NVIDIA DreamDojo (arXiv 2602.06949), built on Cosmos-Predict2.5. NVIDIA Open Model License. Attribution to NVIDIA is required.