Enabling robots to understand and anticipate human behavior
My work in human-robot interaction focuses on enabling robots to perceive, predict, and respond to human intent. The application is not limited to industrial settings, but also healthcare and rehab.
Human activity recognition for effective rehabilitation (Ongoing) Human activity recognition is a crucial capability for effective, AI-assisted rehabilitation. This ongoing project explores learning-based perception methods to understand and anticipate human motion, with the goal of enabling more personalized and effective rehabilitation techniques. The work is being developed in collaboration with a multidisciplinary team of health professionals and artists. Decentralized Perception for Human Intent Prediction — IEEE RA-L | IROS 2025
Predicting human actions in industrial environments is essential for safe and effective human-robot collaboration. In practice, robots often operate with partial views, occlusions, and sensor noise, making reliable intent inference difficult from a single perspective. This work introduces a decentralized shared-perception framework for mobile robots. Each robot builds a spatial graph of the human and surrounding objects, shares compact information with its neighbors, and fuses these observations over time to infer the human's ongoing action. A swarm-inspired consensus mechanism then aligns the robots' predictions into a unified team-level interpretation, improving robustness in dynamic settings. Key contributions: - Spatial understanding via graph neural networks to model human-object relations. - Temporal understanding that aggregates information from multiple robots in a decentralized manner. - A consensus mechanism that helps predictions converge across the team, increasing resilience. - Empirical results showing improved accuracy with more robots and longer temporal context.
Explore this page to dive deeper into the workings of this model. Read the paper here. For an overview, see the video below (it may take a while to load).
