Dongkkka/groot17_dashboard_0904_10k_32bs

LeRobot GR00T N1.7 — final 10,000-step checkpoint

Batch 32, action chunk 16, three camera inputs, 22-dimensional state/actions. Full trained weights and saved image preprocessing and normalization are included. Real-robot task success has not been evaluated.

Load with standard LeRobot

Use LeRobot 0.6.1 with its GR00T dependencies. No custom loader or engine patch is required. The standard base_model_path automatically downloads the public native dependency to initialize the same 16-step architecture before loading the full policy weights. The first load therefore also downloads that dependency's weights. Tokenizer/image assets are resolved from nvidia/Cosmos-Reason2-2B.

from lerobot.policies.groot.modeling_groot import GrootPolicy
from lerobot.policies.factory import make_pre_post_processors

repo_id = "Dongkkka/groot17_dashboard_0904_10k_32bs"
policy = GrootPolicy.from_pretrained(repo_id).to("cuda").eval()
preprocessor, postprocessor = make_pre_post_processors(
    policy_cfg=policy.config,
    pretrained_path=repo_id,
    preprocessor_overrides={"device_processor": {"device": "cuda"}},
)

For Cyclo's existing LeRobot engine, download this repository and load its local snapshot directory through the usual model selection flow. Use all three camera keys specified in config.json, a 22-dimensional state, and task text. Always use the saved preprocessor and postprocessor. The first 19 action channels are absolute position commands; the last three are base velocity commands. No relative-action conversion is used.