---
configs:
- config_name: results
  data_files:
  - split: test
    path: evaluation_summary.jsonl
tags:
- robotics
- human-robot-interaction
- evaluation
---

# HRI-diff large: HITBench closed-loop evaluation

Model: https://huggingface.co/nhatcm/hri_diff_large

This release contains 180 object-perturbed tomato-to-bowl rollouts
across nine interactions, with 20 frozen states per interaction. Each rollout
contains a JSON summary, trajectory arrays, and an MP4. The browsable
`evaluation_summary.jsonl` gives relative paths to those files.

The evaluation used the sealed perturbation bank and human-plan initialization,
50-action prediction, one-action execution with replanning, a 650-step cap,
and raw policy actions without a safety filter or transfer guards.

Successes: overall 6/180, task
4/180, HRI 84/180,
safety 166/180.