GR00T N1.7 checkpoints for all 30 subtasks and 10 long-horizon tasks (per-episode instruction protocol)
This PR adds the evaluated GR00T N1.7 checkpoints to ckpt/gr00t-n1.7-multi-instruction/:
| Path | What | Count | Size |
|---|---|---|---|
ckpt/gr00t-n1.7-multi-instruction/<Task>/ |
one checkpoint per subtask, trained under the per-episode instruction protocol: each of the task's 100 episodes carries one of the task's four registered phrasings (25 each), and evaluation draws a phrasing per episode from the same list | 30 | 30 × 11.7 GB |
ckpt/gr00t-n1.7-multi-instruction/long-horizon/<Task>/ |
one checkpoint per long-horizon task, trained the same way on the union of its stages' datasets | 10 | 10 × 11.7 GB |
641 files, ~470 GB. Nothing outside ckpt/gr00t-n1.7-multi-instruction/ is touched.
Naming
Task directories use the names of the paper. Subtasks (Table 2): OpenCabinet, PickBottle, UnscrewCap, PourBeads, OpenDrawer, PickCharger, InsertPlug, PickBolt, SeatNut, TurnNut, OpenLaptop, TypeKeys, ClickMouse, LayFlatPanel, InsertRail, FastenScrew, AspirateLiquid, DispenseLiquid, OpenFridge, AppleToBlender, CloseFridge, PlaceLid, PressButton, OpenPackage, ExtractApple, PickDebris, DumpDustpan, OpenLunchBox, PlaceBread, ScoopTomato. Long-horizon tasks (Figure 2, without spaces): MealPreparation, CabinetAndPouring, DrawerAndPlug, NutAndBolt, OfficeDesk, FurnitureAssembly, PipetteTransfer, JuicePreparation, PackageOpening, TableCleaning. ckpt/gr00t-n1.7-multi-instruction/README.md maps every name to its environment id and to the training run it came from.
What each directory holds
The final checkpoint of the run (checkpoint-12000): weight shards model-0000{1,2,3}-of-00003.safetensors with model.safetensors.index.json, config.json, processor_config.json, statistics.json (normalisation), embodiment_id.json, experiment_cfg/ (the exact training configuration), trainer_state.json (loss curve) and training_args.bin. Optimizer state is not included. The backbone is referenced by its hub id nvidia/Cosmos-Reason2-2B (gated) in config.json and processor_config.json; to load it from a local copy instead, point GROOT_BACKBONE_PATH at that copy (algo/groot.py in the DexArena repository rewrites the reference at load time).
Recipe
GR00T N1.7 full-parameter finetune (vision and language backbone unfrozen), global batch 2 × 16 gradient accumulation, learning rate 4e-5, 12 000 steps, one A100-80G, 100 teleoperated episodes per subtask (the data/sharpa_hand/ release of 2026-09-15, relabelled per episode as described above). Observations: a fixed base camera plus the wrist camera(s) at 256 × 256; state and action are controller-ordered joint targets (29 for one arm + hand, 58 for the dual-arm OpenPackage).
Evaluate with the DexArena repository: scripts/dex/eval_groot.sh <ENV_ID> <checkpoint dir> (50 episodes, seeds 10001–10050, action horizon 8, instruction drawn per episode from the task's registered phrasings).
Verification
Every file was compared with the staged original after upload (scripts/dex/ops/upload_hf_checkpoints.py --check --pr): 641 files present, none missing, no size differs. Uploaded through this PR because the uploading account is a contributor of the organisation and cannot push to main; merging needs a member with write access.
Upload complete and verified (641 + 641 files); ready for review and merge.