You need to agree to share your Hugging Face information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this model content.

K1 person-following GR00T N1.7

GR00T N1.7 3B fine-tuned to make a Booster K1 humanoid follow a person named by shirt colour ("follow the person in the red shirt") in Isaac Sim, through low clutter and past other people.

GR00T modality key content
video ego_view K1 head camera, 488 × 448 (H × W), ≈95° × 99° FOV, 10 Hz
state base_velocity measured (vx, vy, wz) in the body frame
action velocity_command 16-step chunks of (vx, vy, wz) at 10 Hz, absolute
language annotation.human.task_description the instruction

Embodiment tag: new_embodiment. The velocity commands are tracked by Booster's k1_walk locomotion policy.

Training

  • Data: 512 simulated episodes (about 196k frames) from a privileged follow teacher. The teacher uses A* around clutter and people, keeps 1.3 m behind the leader with the leader in view, and, if the leader starts out of view, searches for it by spinning counter-clockwise. 1–4 animated people are in each episode, with shirt and trouser colours randomized. On 20% of steps a perturbed command was executed while the clean teacher command was recorded, so the data includes recoveries.
  • Tuned: the projector, the diffusion action head and the VL layer norm (tune_projector, tune_diffusion_model, tune_vlln). The language model and vision encoder are frozen.
  • Schedule: 20,000 steps at batch size 32, learning rate 1e-4 with 5% warmup, weight decay 1e-5.

Use

The simulation environment and closed-loop rollout scripts are in the companion repository k1-gr00t-isaac-sim. Its README covers setup; download this checkpoint into it with

hf download wcong521/k1-follow-gr00t --local-dir checkpoints/k1-follow-gr00t

The model loads its vision-language backbone from nvidia/Cosmos-Reason2-2B (gated: accept its license first). To run on a CPU, set use_flash_attention to false in config.json; the repository's policy server does this automatically for --device cpu.

License

Derived from GR00T N1.7 and distributed under the NVIDIA License included in LICENSE: non-commercial use only (research or evaluation purposes).

Downloads last month
-
Safetensors
Model size
3B params
Tensor type
F32
·
Video Preview
loading

Model tree for wcong521/k1-follow-gr00t

Finetuned
(289)
this model