Research

I develop end-to-end robotics systems that bridge the gap between AI capabilities and deployment, encompassing perception, task specification, control, and evaluation under real-world constraints (including latency, safety, partial observability, and imperfect calibration).

My focus is on robust grasping and manipulation of unknown objects in clutter by combining vision with proprioceptive feedback and practical sensing, aiming for performance that does not rely on fragile assumptions.

I study interaction where humans and robots share space, timing, and intent. This includes designing interfaces and behaviors that are predictable, safe, and efficient for users, especially in tasks that require continuous adjustment rather than pre-scripted motions.

I design teleoperation pipelines that enable a human operator to provide intent, while the robot handles timing, collision avoidance, and local motion generation. The goal is reliable task execution, not just demonstrations.

I use wearable sensing, especially EMG, to infer user intent for robot control and assistance. This includes signal processing, intent decoding, and control interfaces that remain usable under day-to-day variability (electrode placement, fatigue, motion artifacts).

I study how vision-language models can support robot locomanipulation by grounding task-relevant language into perception and action in real-world environments. This includes identifying objects and affordances in clutter, guiding shared autonomy, and supporting manipulation on mobile platforms under partial observability, latency constraints, and changing task conditions.

I develop assistive human-robot interaction methods that translate human intent into reliable robot assistance through adaptive interfaces, shared autonomy, and real-time perception. This includes collaborative work on assistive quadruped robots for mobility support, where onboard sensing and intention-aware control are used to provide context-sensitive assistance in everyday environments. I am also interested in how vision-language models can enhance these systems by improving scene interpretation, user adaptation, and natural interaction with robotic assistants.

I integrate multiple sensing streams (vision, motion capture when available, robot proprioception, and task context) to support manipulation pipelines that are resilient to occlusion, clutter, and changing lighting.

I extend manipulation to mobile platforms (e.g., legged robots) where navigation and manipulation must be solved in the same loop. This requires whole-body planning and control, as well as interfaces that maintain system stability while interacting with the environment.

In the long term, I envision robots that continually improve through their own experience, collecting data from interactions, detecting failure modes, and adapting to new environments with minimal human intervention. The emphasis is on safety, repeatability, and measurable gains over time.