pi05_wta_robopro_25k

pi0.5 with a winner-take-all (WTA) multimodal action head, meta-trained on the full RoboPro roboreal_lerobot multitask dataset (15,999 episodes / 1,622 tasks).

• Base: warm-started from mzxuan/robopro_jax_30000 (pi05_robopro_top_cam_jax).
• WTA head: K=2 latent modes, mode_latent_dim 16, wta_eps 0.05.
• Frozen VLM: SigLIP + prefix Gemma frozen; only the action expert, projections, and mode head trained.
• Recipe: matches jax_30000 (cosine, batch 192); 25k-step checkpoint — the highest from the fully uninterrupted pre-resume run (a step-30k OOM forced a resume; loss was flat ~0.001 across 25k–30k).
• Camera: cam_high = overhead countertop view.

openpi config pi05_jax30000_wta. Holds params/ + assets/ (roboreal_lerobot norm stats); optimizer state omitted.