DecentVLA/pi05_cubestack_fl_client1

pi0.5 CubeStack -- FedAvg per-client final model (c1)

Final-round local model of client 1 -- manuka: Blue_on_Green, Blue_on_Orange, Orange_on_Green from the 2-client FedAvg run (examples/pi05_cubestack_fl_explicit.yaml, 30 rounds x 250 local steps). This is the federated client's weights after its last local update -- the round-29 client state, which was aggregated into the round-30 global model DecentVLA/pi05_cubestack_fl_2client.

What this is / is not:

• IS: the global FedAvg model adapted to this client's data for 250 local steps in the final round (uses the SHARED pooled-6-repo normalization).
• IS NOT: the from-scratch local-only baseline DecentVLA/pi05_cubestack_local_c1 (that one only ever saw this client's 3 repos, its own 3-repo normalization).

lerobot-native export (portable tokenizer, real 6-D action/state). Deploy like the other pi0.5 CubeStack models (SO-101, cameras base_0_rgb/left_wrist_0_rgb, --policy.compile_model=false).

Trained on Isambard-AI (GH200) with decent-vla.