Blogs & Notes
[Post-training: Robot Policy Fine-tuning] Fine-tuning robot policies: decisions to make before training
A checklist covering action formats, normalization, temporal alignment, loss masking, and training budget. Uses ABC-130K with pi0.5, DreamZero, and OpenWAM-α as running examples.
[Post-training: Video Model Distillation] DMD training notes for video model distillation
An implementation walkthrough of DMD and self-forcing, covering teacher and critic roles, the pseudo-MSE objective, gradient flow, and rolling KV caches.