nhatcm/hri_diff_large

HRI-diff large, H50, 100k steps

This is the 768-wide HRI-diff joint checkpoint trained for 100,000 steps on HITBench tomato-to-bowl. model.pt includes the conditioning encoder, trust flow, action decoder, frozen trust-tube tokenizer, and latent statistics.

The policy takes RGB-D at causal offsets [-9, -6, -3, 0], current 8-D robot state, and separate regular and interaction prompts. It predicts 50 continuous actions. In the closed-loop benchmark it executes one action and replans at every step using 20 flow integration steps.

Evaluation artifacts: https://huggingface.co/datasets/nhatcm/hri_diff_large_eval

from huggingface_hub import snapshot_download
from hitbench.hitbench.models.hri_diff.training import HRIDiffPhaseTwo

path = snapshot_download("nhatcm/hri_diff_large")
model = HRIDiffPhaseTwo.from_pretrained(path, map_location="cpu")

The flow-matching dependency is licensed CC BY-NC; check its terms before commercial use.