task0003-hierarchical-groot-n1d7-80k

Task0003 Hierarchical GR00T N1.7 (80k)

This is a full, merged inference checkpoint. It combines the temporal F2 GR00T base, the five-class phase predictor, and the 80,000-step jointly trained action head.

The model expects:

• head-camera RGB frames at offsets [-8, -4, 0];
• current arm_left and arm_right state, eight values per arm;
• the exact full instruction: Grasp the trash and drop it into the left trash bin, then grasp the plate and put it into the right blue bin.

It predicts 30 action steps for both arms. Execute 5 steps by default, then observe and infer again. Flush any buffered actions when the stabilized phase changes.

Standard GR00T policy interface

Install Isaac-GR00T N1.7 and make this downloaded model directory importable. The included runtime hook creates the phase-prediction modules before Transformers loads the merged weights.

import sys
from pathlib import Path

model_dir = Path("/path/to/downloaded/model")
sys.path.insert(0, str(model_dir))

from hierarchical_runtime import enable_auto_install
enable_auto_install()

from gr00t.policy.gr00t_policy import Gr00tPolicy

policy = Gr00tPolicy(
    embodiment_tag="NEW_EMBODIMENT",
    model_path=str(model_dir),
    device="cuda:0",
)
actions, info = policy.get_action(observation)

Phase-aware policy interface

Use this interface when phase probabilities, stable phase names, buffer-flush signals, and the recommended execution horizon are required.

import sys
from pathlib import Path

model_dir = Path("/path/to/downloaded/model")
sys.path.insert(0, str(model_dir))

from hierarchical_policy import FullHierarchicalGr00tPolicy

policy = FullHierarchicalGr00tPolicy(
    model_path=model_dir,
    device="cuda:0",
)
actions, phase_info = policy.get_action(observation)

The underlying VLM dependency is nvidia/Cosmos-Reason2-2B.

Validation

• merged step: 80,000 / 80,000;
• all eight phase-related tensors exactly match the trained joint delta;
• single-directory CPU loading passed;
• B200 forward inference passed;
• standard Gr00tPolicy compatibility passed;
• output shapes: arm_left=(1,30,8), arm_right=(1,30,8);
• phase output: five normalized class probabilities.