You need to agree to share your contact information to access this model
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This model is a Derivative Model of NVIDIA GR00T-N1.7-3B and embeds NVIDIA Cosmos-Reason2-2B weights. It is distributed under the NVIDIA Open Model License Agreement. By requesting access you agree to that Agreement (https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license/), including its redistribution and attribution terms. Access is gated to match the upstream Cosmos-Reason2-2B posture.
Log in or Sign Up to review the conditions and access this model content.
GR00T-N1.7 Unitree-G1 — Apple-to-Plate (whole-body loco-manipulation)
A GR00T N1.7 Unitree-G1 adapter, fine-tuned for the whole-body
loco-manipulation task "pick up the apple, walk left and place the apple on the
plate" and exported as a full merged 3B checkpoint that loads directly with
gr00t/eval/run_gr00t_server.py --embodiment-tag UNITREE_G1.
Built on NVIDIA Cosmos. Licensed by NVIDIA Corporation under the NVIDIA Open Model License. See
NOTICEandLICENSE.
| Success — apple picked, walked left, placed on plate | Failure — grasp + walk OK, placement misses |
|---|---|
 |
 |
What this is
NVIDIA published no N1.7 Unitree-G1 checkpoint (UNITREE_G1 is a post-train
embodiment tag — the base model gives ~0% on it without fine-tuning). This
adapter fills that gap so the G1 whole-body loco-manipulation task can be run on
the GR00T N1.7 stack (Cosmos-Reason2-2B backbone), not just N1.6.
- Base model:
nvidia/GR00T-N1.7-3B(VLM backbone = Cosmos-Reason2-2B) - Embodiment:
UNITREE_G1(unitree_g1_full_body_with_waist_height_nav_cmd) - Task / sim:
LMPnPAppleToPlateDC_G1_gear_wbc(NVlabs GR00T-WholeBodyControl, GEAR WBC + MuJoCo) - Fine-tune mode: frozen backbone; only the action projector + flow-matching (diffusion) action head are trained
- action_horizon: 50
Evaluation
On the LMPnPAppleToPlateDC_G1_gear_wbc rollout (10 episodes, n_envs=1,
g6.12xlarge), this checkpoint scored success_rate = 0.7 (7/10). It was
trained by behavioral cloning on demonstrations collected from an N1.6 teacher
(a community re-finetune), so it is a capability/behavior demo rather than a
performance-tuned policy; your numbers will depend on your own fine-tune and
evaluation protocol.
⚠️ action_horizon consistency (important if you re-export)
The exported model config (config.json + experiment_cfg/final_model_config.json
action_horizon) must match the processor's action delta_indices length
(50 here). A mismatch (e.g. model 40 vs processor 50) is silent at export but
crashes a rollout at step 0 with
operands could not be broadcast together with shapes (1,40,D) (50,D). The
weights are horizon-agnostic, so fixing this is a no-retrain edit (set the config
action_horizon to match the processor and re-save).
How to run
This is a full merged HF checkpoint. Serve it with the GR00T inference server:
python gr00t/eval/run_gr00t_server.py \
--model-path <this-checkpoint> \
--embodiment-tag UNITREE_G1 \
--use-sim-policy-wrapper
The Cosmos-Reason2-2B backbone is re-fetched at load time, so the serving host
needs Hugging Face access to the gated nvidia/Cosmos-Reason2-2B repo.
A one-command AWS deployment that pairs this checkpoint with the
GR00T-WholeBodyControl sim is available in
sample-vla-simulator-on-aws
(target gr00t-g1).
License
NVIDIA Open Model License Agreement. This Derivative Model embeds
nvidia/Cosmos-Reason2-2B weights (frozen, unmodified) via the
nvidia/GR00T-N1.7-3B base; redistribution follows Section 3 of that Agreement
(copy of the Agreement + NOTICE attribution + "Built on NVIDIA Cosmos"). Access
is gated to match the upstream Cosmos-Reason2-2B posture.
- Downloads last month
- 24
Model tree for andrewc76/GR00T-N1.7-G1-AppleToPlate
Base model
nvidia/GR00T-N1.7-3B