Pengcheng Xi

I am a Senior Research Scientist at the National Research Council Canada (NRC) in Ottawa, where I lead projects at the intersection of AI, digital health, human motion, and robotics.

I also serve as an Adjunct Research Professor at Carleton University. Previously, I held an Adjunct Assistant Professor position at the University of Waterloo.

I received my PhD in Electrical and Computer Engineering from Carleton University, where I was co-advised by Rafik Goubran and Chang Shu.

Email  /  Scholar  /  X  /  LinkedIn

Research

My research spans machine learning, robotics, computer vision, and human-centered systems. I focus on human–autonomy teaming, studying human trust, situational awareness, and collaboration frameworks that enable people and intelligent systems to work together effectively.

I work on human–robot interaction in assistive contexts, particularly for older adults, with emphasis on physical collaboration and human perception. This aligns with my research in 3D human motion generation, where we develop AI methods for realistic movement synthesis to support motion understanding and interaction design.

In AI for digital health, I build multimodal sensing and analytics systems using audio, video, physiological signals, and data fusion for safety, wellbeing, and early detection. I also study diet and food understanding by integrating visual and linguistic cues with unsupervised learning and generative AI.

Earlier work in 3D human modeling and anthropometric analysis informs my current efforts to advance human-centered AI and assistive technologies.

Updates

• Oct 2025: Paper accepted to SIGGRAPH Asia 2025 (Technical Communications).
• Sept 2025: Paper accepted to HUMANOIDS 2025 (ARC workshop).
• Sept 2025: Paper accepted to CoRL 2025 (Data in Robotics Workshop).
• Sept 2025: Paper accepted to SIGGRAPH Asia 2025.
• July 2025: Invited to review for Croatian Science Foundation (HRZZ).
• April 2025: Paper accepted to IEEE Transactions on Instrumentation and Measurement.
• January 2025: Started new project on generative motion models.

SHARE is a new method for accurately placing 3D human motion into real scenes from a single monocular video. By aligning human meshes with scene geometry and ensuring temporal consistency, SHARE produces more realistic human–environment interactions and outperforms existing reconstruction approaches on both curated and in-the-wild videos.

We introduce StableMotion, a diffusion-based approach for cleaning corrupted motion capture data without the need for paired clean datasets. By leveraging simple motion quality indicators, our method trains directly on raw, artifact-prone mocap data and can both detect and fix errors at test time. Applied to a 245-hour SoccerMocap dataset, StableMotion reduces motion pops by 68% and frozen frames by 81%, outperforming state-of-the-art methods while preserving motion content.

We investigate how different teleoperation interfaces shape the quality of robot learning data. In a study of 400 demonstrations on assistive tasks (wiping and lamp switching), we compare VR controllers and a haptic device, measuring operator workload with NASA-TLX and fine-tuning the Octo model under varying conditions. Our findings reveal clear modality-driven differences in data quality and downstream imitation learning performance, offering practical guidance for designing robot data collection strategies.

We present a comparative study on how different teleoperation modalities—markerless hand-tracking, VR controllers, and haptic devices—affect the quality of demonstrations and learned policies in imitation learning. Focusing on assistive manipulation tasks relevant to aging populations, we show that VR and haptic inputs provide more consistent demonstrations than hand-tracking alone. Moreover, combining modalities improves policy robustness and accuracy.

This paper presents a lightweight, embedded cough analysis model for smart home systems, enabling accurate and privacy-preserving health monitoring. By distilling knowledge from a large speech foundation model, the student network achieves strong performance across multiple cough-related tasks while dramatically reducing model size and computation. The proposed approach generalizes well to unseen sound classes and eliminates the need for cloud-based audio processing.

The paper presents an approach to modeling food degradation using fuzzy clustering, capturing the gradual nature of decay without relying on labeled data. It integrates traditional visual features with semantic features from Vision-Language Models to enable both low-level and high-level interpretation.