Current Research

Research Area 1: Environment-aware dynamic safety for multiagent systems 

State-of-the-art literature on constrained multiagent system operations can only deal with constant or at best time-varying constraint requirements. Such constraint formulations cannot respond well to the dynamic environment. Furthermore, in practical scenarios, external agents outside of the multiagent systems can also influence the safety constraint requirements. Consider a real-world example from the nature, where a fish school faces attacksfrom predators. We can observe that each fish will attempt to swim as close to other fish as possible while avoiding collisions; whereas in the absence of predators, the inter-fish distances can be relaxed inside the fish school. Meanwhile, to ensure survival, each fish should not separate too far from therest of the school. Similarly, when an “attacker” is approaching an autonomous multiagent system, the safety and performance constraints should adapt dynamically, with the desired path and desired path speed also change accordingly. This requires the safety constraint should be dynamic in nature, which depend on the presence of external agents.

Due to the complex operating environment for the multiagent systems, safety and performance constraintrequirements cannot be merely constant or time-varying, but instead need to depend on path parameters and the attacker. The multiagent desired paths and desired path speeds are also path and attacker-dependent. We aim to propose learning-based cooperative control architectures to address environment-aware dynamic safety constraint requirements. The proposed architectures will ensure exponential convergence of tracking errors, while at the same time guarantee safety of the team.

Research Area 2: Resilient cooperative load carrying by a team of UAVs under multiple operation constraints

Using a team of unmanned aerial vehicles (UAVs) to carry a common load has been widely studied in the literature. However, most works assumed that the load to be carried is dimensionless and/or rigid. Some also assume that the attach point is at the load center of the mass. Such assumptions are usually not representing the real-world operating conditions. Furthermore, during such cooperative operation, the UAV team must simultaneously satisfy multiple operation constraints on the inter-agent distance and UAV attitude, as well as speed and/or acceleration. Such considerations further complicate the analysis and implementation of the cooperative control schemes.

This work examines the development of adaptive constrained learning architectures for multiagent systems, in the particular context of cooperative load carrying by a team of UAVs, where we do not assume the carried load is rigid or dimensionless. Adaptive laws will be used to handle not only the system unknown parameters, but also environment uncertainties. Realistic operation constraints must also be considered and integrated into the developed architectures, to ensure the safe, precise, and robust operation of the team. Moreover, such architectures will also consider the resiliency issues such as malfunctioning of some agents, so that to continue carry out the task when undesirable interruptions occur.

Research Area 3: Adaptive iterative learning control under non-uniform learning conditions

Since the early work by Arimoto, ILC algorithms have been well developed and are known to be effective in handling repetitive control processes. ILC has been widely used in industrial applications, especially for robotic systems and motion control systems. However, most ILC algorithms in the literature require that the operation time length for each iteration to be strictly identical. That is, the operation time for each round of task should be exactly the same, and the learning process can only occur over these trials with identical length. However, this is a very restrictive assumption for the operation of autonomous vehicles, since any minor change of operating conditions can end up with a different operation time for each iteration. Currently there are very few works in the ILC literature that have considered the problems of iteration varying trial lengths, with some few exceptions but under other restrictive operating conditions.

We aim to build new adaptive iterative learning cooperative algorithms that are capable of both learning over the iteration domain, which is the feature of iterative learning control, and adaptation over the time domain, which is the feature of adaptive control. This new architecture in the context of multi-vehicle formation operation is named “Adaptive Iterative Learning Cooperative Control (AILCC)” architecture. The AILCC architecture proposed here is fundamentally different from the notion of “adaptive iterative learning control” (AILC) in the literature, where the “adaptive” part of the controller is to update the parameter estimations iteratively, instead of updating the control signal itself directly. It means that the learning is only happening over the iteration domain instead of the time domain. Therefore such AILC frameworks cannot address time-varying non-repetitive nonlinearities and uncertainties.

Research Area 4: Constrained formation control of a tractor-trailer vehilce team

Autonomous articulated vehicles, or tractor-trailer-type vehicles, have received growing attention recently, due to their modular design and larger payload capacity. However, current works mainly focus on the use of a single articulated vehicle. No work has investigated generic cooperative tasks for a team of autonomous articulated vehicles, which would have wide range of applications in logistic and transportation tasks at warehouse yards, distribution centers, and container terminals, such as teams of articulated buses, trucks, and guided vehicles. Although there exists tons of literature on cooperative systems, such algorithms cannot be applied to articulated vehicle teams, due to their special and highly underactuated configurations.

This project focuses on the safe, precise, and robust cooperative formation operations of an articulated vehicle team. Specifically, safety constraints need to ensure collision avoidance between any two points of neighboring articulated vehicles, and precision constraints require precise tracking of a desired formation pattern. However, current cooperative control algorithms on non-articulated vehicle teams only focus on constant or time-varying constraint requirements, which often require more aggressive kinematic behavior that can eventually cause actuation saturation. The geometric and spatial nature of the constraint requirements is often overlooked, largely due to difficulties in integrating time-domain system kinematics and control with pathdomain constraints. Furthermore, the highly underactuated nature of articulated vehicles makes the operation subject to multiple feasibility constraints, which can result in infeasible control actions if violated.