Publication
HINT-Bench: human intention recognition benchmark for social robotics
Conference Article
Conference
Iberian Robotics Conference (ROBOT)
Edition
2025
Doc link
File
Abstract
In modern indoor environments such as hospitals, offices, and homes, service robots must move beyond reactive behaviors and anticipate user needs by inferring human intent. Early intention recognition enables proactive assistance, thereby enhancing efficiency, safety, and user experience. We present an open‐source benchmark suite for early human intention recognition that integrates (1) a high‐fidelity Gazebo simulation with ROS 1, featuring three Soft Actor–Critic (SAC)‐trained agents modeling collaborative, neutral, and adversarial behaviors; (2) multimodal perception comprising 9D LiDAR/odometry state vectors and 135D MediaPipe skeleton keypoints; and (3) two curated datasets: a 300‐episode training set and a 300‐episode test set pre‐sliced into 500 spatial (1–5 m) and 500 temporal (1–9 s) trigger snapshots per class.
We benchmark six baseline methods, including approaches based on trajectory or skeleton data.
Our unified evaluation toolkit computes accuracy, precision, recall, F1 score, mean time-to-correct-prediction, noise robustness, and inference latency (CPU/GPU). All code, data, and scripts are available at https://github.com/valerio-bo/HINT-Bench, offering a reproducible platform to accelerate research in anticipative human–robot collaboration.
Categories
humanoid robots, learning (artificial intelligence).
Author keywords
intent prediction, human-robot interaction, social robotics
Scientific reference
V. Bo, A. Garrell Zulueta and A. Sanfeliu. HINT-Bench: human intention recognition benchmark for social robotics, 2025 Iberian Robotics Conference, 2025, Porto, Portugal, Springer, to appear.
Follow us!