Incline_new_60epi_GR00T17

GR00T N1.7 finetuned on a bimanual OpenArm ramp-loading dataset. The run went the full 8,000 steps and ended on its own; nothing here was cut short.

Load

The default branch holds the final checkpoint, 008000:

--policy.path=leapshared/Incline_new_60epi_GR00T17

Each earlier checkpoint is a branch of the same repo, with the weights at the branch root:

--policy.path=leapshared/Incline_new_60epi_GR00T17 --policy.pretrained_revision=004000

--policy.path takes a repo id or a local directory and reads config.json from the root only — it has no subfolder syntax, which is why these are branches rather than a checkpoints/ folder.

Only the weights and the pre/post-processors are here. Optimizer state was not uploaded, so this repo serves inference, not resumption.

Checkpoints

Checkpoint	Epochs	Train loss	Revision to pass
002000	0.48	0.059	--policy.pretrained_revision=002000
004000	0.96	0.039	--policy.pretrained_revision=004000
006000	1.44	0.025	--policy.pretrained_revision=006000
008000	1.92	0.021	default branch, no revision needed

Training

Base model	nvidia/GR00T-N1.7-3B
Policy type	groot_frozen_bf16 (frozen backbone held in bf16, fused AdamW)
Dataset	leapshared/Incline_new_20260902_203009 — 60 episodes, 66,823 frames, 30 fps, bi_openarm_follower
Cameras	follower_d455f, left_wrist, right_wrist
State / action	16 dimensions, absolute (use_relative_actions=false)
Chunking	chunk_size=40, n_action_steps=40
Trained modules	projector, diffusion model, VL layernorm. LLM and vision tower frozen
Optimizer	fused AdamW, lr 1e-4, weight decay 1e-5, grad clip 1.0
Schedule	cosine, 500 warmup steps, 8,000 total
Batch size	16
Seed	42
Image transforms	enabled

Measured on one 32 GB card: 28.29 GiB resident, 2.79 step/s, 48 minutes 57 seconds end to end. Final step logged loss:0.021 grdn:0.200.

Task

Place the four metal weights on the blue cart on the ramp.