Downloads · 30 days
8
36% of all-time downloads
gr00t-n15-robocasa-base60k
mlnha/gr00t-n15-robocasa-base60k
gr00t-n15-robocasa-base60k is a robotics model from mlnha. Use it for the robotics task on the model card, and read the license before you ship it in a product. The card lists the license as other.
The base checkpoint every augmentation experiment in the pose6daug (NeurIPS) RoboCasa lineage starts from. All four paper conditions — Base, Ours/Pose6DAug, VACE, MimicGen — resume from this and train a further 30k st…
All-time downloads
22
Public
Parameters
2.7B
7.6 GB on disk
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors7.6 GB · 100%
Parameter types
How the weights are stored.
BF161.7B · 61%
Model card
From the Hugging Face model README
GR00T-N1.5-3B fine-tuned on RoboCasa PickPlaceCounterToCabinet — 60k steps
The base checkpoint every augmentation experiment in the pose6daug (NeurIPS) RoboCasa lineage starts from. All four paper conditions — Base, Ours/Pose6DAug, VACE, MimicGen — resume from this and train a further 30k steps on their own dataset.
| base model | nvidia/GR00T-N1.5-3B |
| data | PickPlaceCounterToCabinet, all 502 human episodes |
| steps | 60,000, global batch 64 (4 GPUs x 16) |
| embodiment tag | new_embodiment |
| data config | single_panda_gripper (RoboCasa Panda, 3 views) |
| backbone | Eagle-2, select-layer 12, LLM and vision tower frozen |
| tuned | projector + diffusion head |
| optimiser | AdamW(0.95, 0.999), wd 1e-5, cosine 1e-4, 5 % warmup, bf16 |
| original run | pnpcountertocab_all502_gbs64_wandb_60k_save5k_20260427_193901 |
Weights only — no optimiser state, so this can seed a fine-tune but not resume the 60k run.
Use
Needs the MoSS/myGR00T fork, not public Isaac-GR00T: the model class and the
single_panda_gripper data config only exist there. The fork travels with the training
bundle (train_robocasa/myGR00T/).
cd train_robocasa
BASE_CKPT=/path/to/this/checkpoint \
DATASET_PATH=/path/to/gr00t_views/<your dataset> \
GPUS=0,1,2,3 MAX_STEPS=30000 SAVE_STEPS=5000 PER_GPU_BATCH=16 \
bash scripts/train_groot15_single_dataset.sh
The launcher checks for model.safetensors.index.json and refuses to start without it.
Contents
model-00001-of-00002.safetensors 4.8 GB
model-00002-of-00002.safetensors 2.5 GB
model.safetensors.index.json
config.json GR00T_N1_MoSS / gr00t_n1_5
experiment_cfg/metadata.json embodiment statistics used for normalisation
trainer_state.json the 60k loss history
experiment_cfg/metadata.json matters: it holds the state/action statistics the transforms
normalise against. A checkpoint without it will train, but against the wrong scale.
- Task
- Robotics
- Type
- gr00t_n1_5
- License
- other
- Created
- Sep 19, 2026
- Updated
- Sep 19, 2026