add README.md

README.md

@@ -0,0 +1,43 @@
+---
+license: apache-2.0
+base_model: nvidia/GR00T-N1.7-3B
+library_name: gr00t
+pipeline_tag: robotics
+tags: [robotics, vla, gr00t, isaac-gr00t, decent-vla, cube-stacking, so-101, federated]
+---
+
+# GR00T N1.7 CubeStack — Centralized (3-client union)
+
+Centralized ceiling: pooled union of all 6 tasks, **37500 steps** (= 3 clients x 50 rounds x 250 local steps, equal compute to the federation).
+
+Trained on Isambard-AI (GH200) with [decent-vla](https://github.com/kevinDuan1/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.