Spring Robotics Colloquium: Chuchu Chen (George Washington University)

Why It Matters

Reliable, inexpensive state estimation unlocks mass‑market AR/VR and autonomous robots, driving safety, privacy compliance, and faster product rollouts.

Key Takeaways

• Visual‑inertial fusion enables cheap, accurate robot state estimation
• Consistency requires covariance bounds not smaller than true uncertainty
• Unobservable directions cause estimator inconsistency in VIO systems
• Observability analysis can restore theoretical guarantees across sensors
• Trustworthy, privacy‑aware robots depend on robust estimation algorithms

Summary

In this colloquium, Assistant Professor Chuchu Chen of George Washington University outlines her visual‑inertial navigation research, arguing that any AI‑enabled device with a physical body— from AR glasses to drones— qualifies as a robot. She stresses that robots must be trustworthy, respect privacy, and operate within resource and dynamical constraints, making reliable state estimation a foundational problem.

Chen explains that cameras provide high‑resolution, inexpensive visual data but lack scale and are sensitive to lighting, while IMUs deliver high‑frequency inertial measurements yet suffer from drift and bias. By modeling both sensors mathematically and treating the fusion as a probabilistic estimation problem, she demonstrates how filtering or sliding‑window optimization can produce estimates of pose, velocity, and orientation along with covariance that quantifies uncertainty.

A key illustration involves AR glasses assisting visually‑impaired shoppers: the system must first locate the user within a GPS‑denied environment and then identify target objects. Chen highlights the concepts of accuracy and consistency, showing that an inconsistent covariance—one that under‑estimates uncertainty—can lead to unsafe robot behavior. She attributes much of this inconsistency to unobservable degrees of freedom in visual‑inertial systems and describes how observability analysis restores consistency with theoretical guarantees across multiple sensor modalities.

The broader implication is that robust, low‑cost visual‑inertial estimation enables trustworthy, privacy‑preserving robots at scale. Industries ranging from consumer AR/VR to autonomous logistics can reduce hardware expenses while improving safety, accelerating commercialization of intelligent embodied AI.

Original Description

Title: Reliable and Efficient Visual-Inertial Estimation and Spatial Perception

Speaker: Chuchu Chen (George Washington University)

Date: Friday, April 17, 2026

Abstract: Visual-inertial systems (VINS), which fuse camera and inertial measurements for motion estimation and spatial perception, have become an important capability for robotics and XR. Due to their low cost, compact size, and complementary sensing modalities, they are particularly attractive for drones, mobile robots, and wearable devices operating under tight sensing, energy, and compute constraints. At the same time, achieving reliable real-time performance on such platforms remains challenging. In this talk, I will present recent work on the foundations of visual-inertial estimation, focusing on estimator consistency, decoupled error and state representations, and efficient visual-inertial odometry for low-cost and low-energy platforms. I will also discuss selected results on multi-sensor calibration and how system design choices affect practical performance in resource-constrained deployment. Finally, I will discuss how these estimation principles connect to spatial perception, including plane-aware estimation and recent efforts toward richer 3D scene representations. Together, these directions support accurate, efficient, and deployable perception for robots operating in complex real-world environments.

Bio: Chuchu Chen is an Assistant Professor in the Department of Mechanical and Aerospace Engineering at The George Washington University, where she directs the EPIC Lab (Estimation, Perception, and Intelligent Computing). She received her Ph.D. in Mechanical Engineering and her M.S. in Computer and Information Sciences from the University of Delaware in 2025, where she was a member of the Robot Perception and Navigation Group advised by Prof. Guoquan (Paul) Huang. Her research focuses on state estimation and spatial perception for robotics and XR, with an emphasis on reliable, efficient, and deployable systems under real-world sensing and compute constraints. Her honors include the ICRA 2024 Best Paper Award Finalist (Robot Vision), the RSS 2023 Best Student Paper Award Finalist, and the University of Delaware Doctoral Fellowship for Excellence.

This video is in the process of being closed captioned.