DecentVLA/smolvla_cubestack_cent_3client

SmolVLA CubeStack — Centralized (3-client union)

Centralized ceiling: full fine-tune on the POOLED union of all 6 tasks (2 per arm across harry/zhekai/kevin), 37500 steps == the federation's total compute budget (3 x 50 rounds x 250 local steps). Shared pooled-6-repo normalizer. Final train loss 0.0074.

lerobot-native export (portable config, real 6-D action/state). SO-101 CubeStack, cameras camera1 (front) / camera2 (wrist), camera3 zero-filled. Trained on Isambard-AI (GH200) with decent-vla. Part of the 3-client SmolVLA CubeStack federated study — SAME color-pair non-IID partition as the pi0.5 3-client study (c0=harry Green/Orange, c1=zhekai Green/Blue, c2=kevin Orange/Blue; each client blind to the 3rd color). TRUE full fine-tune (unfrozen VLM, no LoRA), SmolVLA LIBERO recipe (lr 1e-4, cosine, batch 32, grad-clip 10).