Downloads · 30 days
14
22% of all-time downloads
leapshared/Incline_50epi_GR00T17
Incline_50epi_GR00T17 is a robotics model from leapshared. Use it for the robotics task on the model card, and read the license before you ship it in a product. It is set up for lerobot. The card lists the license as apache-2.0.
GR00T N1.7 finetuned on a bimanual OpenArm incline-manipulation dataset.
Downloads · 30 days
14
22% of all-time downloads
All-time downloads
64
Public
Parameters
3.1B
38.1 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors9.5 GB · 100%
How the weights are stored.
F321.6B · 52%
From the Hugging Face model README
GR00T N1.7 finetuned on a bimanual OpenArm incline-manipulation dataset.
The default branch holds checkpoint 041476, the last checkpoint of the run. Earlier checkpoints are on branches named after their step, for step-to-step comparison. Only the model weights and the pre/post-processors were uploaded; optimizer state was not, so this repo serves inference, not resumption.
The default branch holds the final checkpoint, 041476:
--policy.path=leapshared/Incline_50epi_GR00T17
Each earlier checkpoint lives on its own branch, with the weights at the branch root:
--policy.path=leapshared/Incline_50epi_GR00T17 --policy.pretrained_revision=020738
--policy.path accepts a repo id or a local directory only — it has no subfolder
syntax, which is why the checkpoints are separate branches rather than subfolders.
| Base model | nvidia/GR00T-N1.7-3B |
| Policy type | groot_frozen_bf16 (frozen backbone held in bf16, fused AdamW) |
| Dataset | leapshared/Incline_20260901_221322 — 50 episodes, 62,210 frames, 30 fps, bi_openarm_follower |
| Cameras | follower_d455f, left_wrist, right_wrist |
| Action space | absolute (use_relative_actions=false), 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 |
| Batch size | 16 |
| Seed | 42 |
| Image transforms | enabled |
The run was planned for 62,210 steps (16 epochs) and was stopped early by the operator at step 43,134 (about 11.1 epochs). Training loss had flattened near 0.010–0.011 and the cosine schedule had decayed the learning rate to 2.3e-5.
| Checkpoint | Epochs | Revision to pass |
|---|---|---|
| 010369 | 2.67 | --policy.pretrained_revision=010369 |
| 020738 | 5.33 | --policy.pretrained_revision=020738 |
| 031107 | 8.00 | --policy.pretrained_revision=031107 |
| 041476 | 10.67 | default branch, no revision needed |
A power loss cut the run at step 12,658; it was resumed from checkpoint 010369, so steps 10,369–12,658 were trained twice against different data ordering. Checkpoints from 020738 onward are unaffected.