HUMAIN: Human-Aware Implicit Social Robot Navigation

How can robots navigate human-populated environments more socially and fluently without requiring expensive, real-time skeletal tracking or complex manual rule-setting? The authors address the gap between the rich social cues humans use (gait, body orientation, pose) and the limited, often oversimplified, sensorimotor inputs typically available to mobile robots.

The authors propose HUMAIN, a two-stage framework that leverages knowledge distillation.

1. Oracle Teacher Training: A high-capacity transformer-based model is trained using privileged, multi-modal data—including RGB images, robot state, and precise 3D skeletal keypoints of surrounding humans. This teacher learns to map these inputs into a rich latent representation that captures complex social dynamics.
2. Sensorimotor Student Training: A lightweight, real-time-ready student model is trained to navigate using only raw RGB images and a navigation goal. The student is trained with a dual-objective: to minimize trajectory prediction error and to align its internal latent features with those of the frozen teacher model. This forces the student to implicitly infer social cues from visual data that it cannot explicitly track.

HUMAIN demonstrates that distilling privileged social knowledge into a lightweight model significantly enhances navigation performance. In extensive experiments, the framework improved trajectory prediction metrics by an average of 29.8% compared to state-of-the-art baselines. By embedding social awareness directly into the planning loop, the robot achieves more human-like, socially compliant navigation without the computational overhead of explicit pose estimation at runtime.

This work bridges the gap between high-level social reasoning and low-level robot control. By demonstrating that robots can learn to 'see' social cues implicitly, the authors provide a pathway for deploying socially aware robots on resource-constrained platforms that lack the sensors or compute power for complex human-tracking pipelines.