My primary research interests lie in the intersection of control theory and machine learning, with applications to connected and autonomous vehicles, robotics, and nonlinear and complex systems.
Some of my ongoing rojects are summarized below.
1. Agricultural Robotic Systems for Efficient Tree Fruit Harvesting (sponsor: USDA, Role: PI)
Fruit harvesting is highly labor-intensive and cost-heavy; Growing labor shortage and rising labor cost have steadily eroded the profitability and sustainability of the fruit industry.
Furthermore, manual picking activities constitute great risks of back strain and musculoskeletal pain to fruit pickers due to repetitive hand motions, awkward postures when picking fruits at high locations
or deep in the canopy, and ascending and descending on ladders with heavy loads. Therefore, there is an imperative need for the development of robotic mass harvesting systems to tackle labor shortage,
lower human injury risks, and improve productivity and profitability of the fruit industry.
In this project, we develop an efficient, practically-viable robotic harvesting system by addressing the following fundamental challenges:
Designing and assembling of multi-arm robot to enable dexterous manipulation and efficient;
Developing accurate and robust perception system to efficiently detect and localize apples as well as tree branches/truncks;
Motion planning and control of the multi-arm robotic system to enable fast, accurate, and robust operation in 3D space with stationary and moving obstacles;
System integration and orchard experiments to evaluate the effectiveness of the system.
2. CAREER: Privacy-Aware Collaborative Sensing and Control for Cloud-Enabled Automotive Vehicles (sponsor: NSF, Role: PI)
With the advent of 5G technology, cloud computing is expected to revolutionize automotive applications by providing “big data” and real-time, high-fidelity computing capabilities. Despite its promise, the use of
cloud computing in automotive vehicle
control and sensing still has limited success due to concerns in communication privacy and real-time constraints inherent to many automotive systems. This project addresses
the major challenges in cloud-based control, collaborative sensing, decentralized optimization, and privacy preservation. The new designs and methodologies will offer a transformative framework
in cloud-facilitated collaborative sensing and control that seamlessly integrate cloud and vehicle resources to enable smarter, safer, and greener next-generation automotive systems.
The cloud-facilitated collaborative sensing and control is expected to greatly enhance vehicle control performance, and achieve improved safety, energy efficiency, and ride comfort.
Four closely integrated research objectives are proposed:
Develop a novel privacy-preserving, learning-based collaborative sensing framework to enable the exploitation of multiple heterogeneous vehicles to iteratively improve the estimation of
important road information (e.g., black ice and pothole) while preserving privacy;
Formalize and synthesize privacy-aware cloud-facilitated control to seamlessly integrate cloud and onboard controls for enhanced performance without leaking vehicle privacy;
Develop a computationally-efficient, privacy-preserving decentralized control framework by explicitly exploiting the sparse communication/constraint topologies in connected vehicles;
Evaluate and validate the proposed frameworks through extensive simulations and experiments.
Soft robots have great promise in versatile applications involving interactions with humans, such as elder care, collaborative surgery, work/life assistance, and collaborative fruit harvesting.
The goal of this project is to develop a novel soft robot system equipped with multiple soft arms,
termed soft multi-arm robot or SMART, and to advance its practical use in applications involving close collaboration with humans.
The soft multi-arm robot system is expected to offer dexterity, efficiency, and intrinsic safety, and achieve productive collaboration with humans with an array of exciting potential applications.
To achieve this goal, five synergistic research thrusts are pursued to overcome key scientific challenges:
Designing and fabricating soft multi-arm robots to realize simultaneous actuation and stiffness-turning and enable dexterous manipulation;
Advancing motion planning and control approaches for these soft robots to achieve robust manipulation in 3D space in the presence of stationary and dynamic obstacles;
Formalizing trust-based human-robot interaction to realize efficient human-robot collaboration by explicitly accommodating the dynamics of human trust in the soft multi-arm robot policy;
Developing orchard and human motion perception algorithms to robustly obtain 3D tree and human position/pose information to support the fruit harvesting application;
Evaluating the soft multi-arm robot system via extensive lab and field experiments in the context of collaborative apple harvesting.
4. Road Information Discovery through Privacy-Preserved Collaborative Estimation in Connected Vehicles (sponsor: NSF, Role: PI)
Real-time and crowd-sourced road information, such as black ice, pothole, and road roughness, can improve vehicle performance. Existing road information discovery approaches are not always practically viable,
due to limitations in road coverage and lack of robustness. We develop a novel collaborative road information crowdsourcing methodology using connected vehicles.
The new methodology will enable efficient, robust, and broad-coverage road information discovery by utilizing connected vehicles as mobile sensors while preserving privacy of the participating vehicles.
The privacy-preserved collaborative estimation using connected vehicles is expected to make vehicle-based road information discovery practically and economically viable. This project will overcome
several scientific challenges to realize full application potential of such connected systems. The main research tasks include:
Developing jump-diffusion process-based estimation to enhance road information discovery performance when dealing with abrupt input/disturbance changes in a single vehicle setting;
Advancing iterative learning-based collaborative estimation across heterogeneous vehicles to enable the exploitation of local estimation from a network of heterogeneous agents to
iteratively improve the performance of road information discovery;
Designing dynamics-enabled privacy preservation schemes to protect vehicle privacy without affecting computation fidelity or incurring large computation/communication overhead;
Evaluating the methodology in the application of collaborative road profile estimation.
5. Learning-empowered Adaptive Modeling and Predictive Control for Automotive Systems (sponsor: Ford, Role: PI)
Automotive systems are highly nonlinear and complex. Identification of its dynamic models and performing model-based control are challenging tasks.
We develop a unified and computational efficient online system identification and model predictive control (MPC) framework to improve vehicle control performance and reduce calibration,
by exploiting a spatio-temporal filtering-based system identification.
From there, we develop a linear parameter varying (LPV) MPC based controller for the obtained multiple linear systems. The final demonstration is planned on vehicle implementation.
The main research tasks include the following:
Developing a spatio-temporal filering (STF) based system identificaiton to systematically decompose a nonlinear sytem into local linear models;
Developing a LPV-MPC framework based on the identified STF models with guaranteed stability;
Evaluating the unified framewok in automotive powertrain testbed.
6. Safe Multi-agent Reinforcement Learning - Theory and Applications
Despite RL's recent success in a broad range of applications such as board
games, natural language processing, and finance, its applications in real-world
engineered systems are still rare. The primary reason is that RL is essentially a trial-and-error
learning process that may violate system constraints during training, which could lead to disastrous
consequences such as battery overheat, robot breakdown, and car crashes. In the meantime,
the dynamical systems and control community have come up with various theories and tools for
predicting and manipulating the evolution of internal system states. In this project, we develop a novel model-enabled SafeRL framework that combines model-free RL with a model based
safety regulator for learning supervision to learn a safe policy safely. The main thrusts of this research include the following:
Safety regulator design: We develop a safety regulator to supervise conventional
learning process by incorporating state-dependent constraint-admissible set which is obtained
with recursive feasibility techniques through exploiting underlying physical dynamics and safetyrelated
system constraints. We consider the safety set for linear system as well as nonlinear uncertain systems.
Regulation mechanism design: We develop schemes to seamlessly integrate the constraint-admissible set in RL's
action exploration to guarantee safety while simultaneously guiding the agent to learn the system
constraints. Specifically, we develop a safety enforcement scheme that projects unsafe exploratory actions to the safe action set. We
adesign a control regulation penalty feedback mechanism to guide the agent
to learn the system constraints. We will perform convergence analysis under a linear value function approximator. This analysis will in return guide the
mechanism design of safety enforcement and regulation feedback.
Safe Multi-Agent RL: We extend the safe RL framework to multi-agent systems to enable safe, efficient, and scalable learning among multiple agents.
Applications: We will apply the Safe Multi-agent RL framework to a number of engineered systems such as mutli-robotic systems, connected and autonomous vehicles, power grids, and automotive engines.
