Research
Robot intelligence for physical systems that work with people.
The Control and Robotics Lab develops planning, control, learning, and prediction methods for robots operating in complex real-world environments. We are especially interested in robots that can anticipate human motion, reason over geometry and contact, adapt safely online, and make sustainable manufacturing workflows more scalable.
Research agenda
We study the full loop from perception and prediction to planning, control, and deployment.
Our research is organized around a practical question: how can robots make reliable decisions when people, contacts, geometry, and hardware constraints are all changing at the same time?
Perceive task state
Segment products, detect components, estimate contacts, and represent scene geometry from vision, point clouds, and force/torque signals.
Anticipate people and dynamics
Model future human motion, infer intent from interaction, and account for uncertainty in collaborative robot behavior.
Generate feasible robot actions
Use graph reasoning, neural warm starts, and optimization-based planners to compute safe manipulator motions in cluttered workspaces.
Close the loop on hardware
Design controllers for manipulators, morphing aerial robots, connected vehicles, and mechatronic systems under uncertainty.
Validate in real systems
Apply the methods to robotic disassembly, remanufacturing, human-robot collaboration, and transformable aerial platforms.
Current directions
Recent work from Prof. Minghui Zheng’s group
Intelligent disassembly
Robot-assisted selective disassembly and remanufacturing
We develop systems that inspect, disassemble, sort, and recover components from end-of-use products. Recent work includes RAISE for end-of-life phones, vision-language-action models for desktop disassembly, and vision-model benchmarking for irregular e-waste component segmentation.
Representative outputs: RAISE, VLA models for selective desktop disassembly, e-waste segmentation benchmarks, ACC 2026 tutorial on electronics remanufacturing.
Human-aware autonomy
Prediction and planning for human-robot collaboration
Our algorithms combine real-time 3D human motion prediction, uncertainty-aware planning, and contact-based intent inference so robots can anticipate people and adapt online.
Representative outputs: Bayesian-optimized diffusion models for real-time 3D human motion prediction, uncertainty-aware graph-based planning, contact-based intent inference.
Learning-based planning
Generalist neural motion planners for manipulators
We study how deep learning can support sampling, steering, collision checking, trajectory optimization, and fast deployment for robot manipulator motion planning.
Representative outputs: T-ASE survey on generalist neural motion planners, SIMPNet, GAIDE, and flow-matching warm starts for optimization-based planning.
Robot foundation models
Safe foundation-model-enabled robots
We investigate modular safety guardrails and vision-language-action interfaces that help high-level robot policies remain constrained, interpretable, and useful in physical tasks.
Representative outputs: modular safety guardrails for foundation-model-enabled robots and VLA policies for disassembly tasks.
Aerial robotics
Morphing aerial robots and dynamic control
Our work on transformable quadrotors studies mechanical design, modeling, and control for robots that adapt their shape while maintaining stable flight in constrained spaces.
Representative output: MorphoCopter, a transformable quad-bi copter with modeling, control, and experimental validation.
Methods
Technical threads across the lab
Learning-based control
Human motion prediction
Robot perception
Optimization
Contact-aware manipulation
Vision-language-action models
Mechatronic system design
Application focus
Robotics for sustainable manufacturing
Disassembly and remanufacturing connect many of the lab’s core research questions: perception under clutter, task and motion planning, safe human-robot collaboration, manipulation under contact, and control of physical systems. This application area gives our methods a demanding testbed with clear societal and industrial value.