DecentVLA/groot_cubestack_fl_3client

GR00T N1.7 CubeStack — FedAvg 3-client (global)

Federated global model: FedAvg over 3 colour-pair clients, 50 rounds x 250 local steps, batch 32. Every round aggregated all 3 clients (updates=3).

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

Scope: action-head-only — the Cosmos-Reason2 backbone is frozen (tune_llm/tune_visual false); the ~1.62B DiT action head plus the embodiment projector are trained. That is the standard NVIDIA new-embodiment recipe: the SO-101 projector slot ships at random init, so tune_projector must stay on.

Embodiment: SO-101 (NEW_EMBODIMENT), state/action 6-D single_arm+gripper, cameras front/wrist. The DecentVLA/Cube_* datasets are LeRobot v3.0, which the GR00T loader cannot read, so they were converted to v2.1 and the video pre-scaled to shortest edge 256 (GR00T resizes to exactly that internally; it cut decode ~3.9x).

Partition: the SAME colour-pair non-IID split as the pi0.5 and SmolVLA CubeStack 3-client studies — c0 harry Green/Orange, c1 zhekai Green/Blue, c2 kevin Orange/Blue, each client blind to the third colour.

Caveat worth knowing: the two kevin prompts are worded differently from the other four ("Put orange cube on the blue cube." vs "Put the orange cube on top of the blue cube."). Kept verbatim so this arm trains on exactly the strings the pi0.5 and SmolVLA arms saw.

Both checkpoints hold the SAME parameter set: 537 tensors, 1.6205 B params, all action_head.* (verified by reading the tensors) — i.e. the frozen-backbone scope.

Checkpoint wrapper differs by arm, because they came off different code paths — load with decent-vla --resume, not from_pretrained:

• FL server_round_0050.pt — {"state": {...}, "round": 49} (server aggregate; round is 0-indexed).
• Centralized round_0075.pt — {"state_dict", "include", "extra", "embodiment", "backend"} (the LocalRunner end-of-run save, which also records embodiment + backend metadata). Note there is no server_round_0075.pt: periodic saves ran every 10 rounds, and 75 is not a multiple of 10, so round 75 exists only as this end-of-run file.

Running these weights (experiment_cfg/)

The checkpoint holds ONLY action_head.* (537 tensors, 1.6205 B). Three things GR00T derives from the training dataset are not in it and not in the public N1.7 base repo, so they are published here under experiment_cfg/:

file	what it is
embodiment_id.json	the projector-slot mapping used in this run
processor_config.json	modality config (state/action index ranges, video keys) + image/state processor settings
statistics.json	state/action normalization statistics

Embodiment slot = 10. new_embodiment is absent from the base repo's embodiment_id.json; the value comes from EMBODIMENT_TAG_TO_PROJECTOR_INDEX in gr00t/model/gr00t_n1d7/processing_gr00t_n1d7.py. Verified against the weights, not just that table: diffing this checkpoint's 14 embodiment-indexed tensors against the base head puts 100.0% of the drift in slot 10 and exactly 0.0 in all 31 other slots (see experiment_cfg/slot_drift.json). Slot 0 is untrained — falling back to it runs without error and produces meaningless actions.

Modality (from processor_config.json): video keys front, wrist; state and action single_arm (0-5) + gripper (5-6); action horizon 16; shortest_image_edge 256.

The training config for this run is included as experiment_cfg/train_config.yaml.

Which statistics for the federated model?

Our GR00T adapter wires normalization per client partition, so each FL client trained against the statistics of its own two repos — there is no single "global" normalizer that training produced. All four are published:

• statistics.json — pooled over all 6 repos (identical to the centralized arm). Use this one for a single-number evaluation of the global model: it is also what the pi0.5 and SmolVLA CubeStack FL models carry, so it keeps the three policies comparable.
• statistics_client{0,1,2}.json — what each client actually trained with (c0 harry Green/Orange, c1 zhekai Green/Blue, c2 kevin Orange/Blue).

They differ substantially — action.gripper.mean is 4.92 / 8.05 / 19.78 per client vs 10.92 pooled — so this choice changes results. Note this is a real difference from the pi0.5 and SmolVLA CubeStack arms, whose lerobot adapter pooled the statistics at training time; disclose it in any cross-policy table.