Overview
Our research seeks to create highly agile robotic systems that can operate with robustness and resiliency in real-world environments. In particular, we investigate methods at the intersection of computational control, computational physics, and machine learning to push robotic systems to their physical limits while also providing guarantees on performance (e.g., safety, task completion). We are pursuing both physical and algorithmic agility (e.g., online planning) as well as robust intelligence, which is often characterized by the careful integration of machine learning and model-based methods. Our research and team spans the Johns Hopkins University Whiting School of Engineering (WSE) and the Johns Hopkins University Applied Physics Laboratory (APL).
Aerobotics
Birds routinely execute impressive maneuvers in the presence of complex aerodynamics to navigate through challenging environments. We are conducting research at the intersection of fluids and robotics to create highly agile aerial robots.
Agile Fixed-Wing Vehicles
Fixed-wing aerial vehicles offer significant performance advantages over rotary-wing vehicles in terms of speed, endurance, and range. However, because fixed-wing vehicles are extremely underactuated and dependent on complex aerodynamics, these vehicles have traditionally been severely limited with regards to maneuverability. To enable extreme fixed-wing agility, we developed a receding-horizon nonlinear model predictive control (NMPC) approach based on direct trajectory optimization that can reason about an expanded flight envelope. We have demonstrated both aggressive maneuvering in constrained environments and precision post-stall landings.
Agile Fixed-Wing Navigation with Onboard Vision
Precision Post-Stall Landing
Controlling the Continuum
Most aerial vehicle flight controllers rely on quasi-steady models that do not consider the complex unsteady effects that often dominate aerobatic flight. We are investigating methods for “sensing and controlling the continuum” by using light-weight computational fluid dynamics in-the-loop for control and estimation. We are also exploring avian-inspired designs, such as dynamic wing morphing.
Intelligent Computational Control
Both computational control and machine learning methods have proven to be powerful tools for controlling high-dimensional, complex robotic systems. However, obtaining performance guarantees for these control systems can be prohibitively challenging, especially in real-world scenarios. We are exploring computational and learning-based control methods that can enable extreme agility while also providing statistical safety guarantees.
Probabilistically-Safe NMPC
To provide performance guarantees in the presence of uncertainty, approaches like robust Nonlinear Model Predictive Control (NMPC) and stochastic NMPC often make limiting assumptions about the form of the dynamics and disturbances. We are investigating approaches for NMPC that can provide performance guarantees for a wide-class of high dimensional stochastic dynamical systems. One of our recent approaches, Probably Approximately Correct (PAC) NMPC, uses recently derived sample complexity bounds to provide anytime statistical performance guarantees for receding-horizon feedback motion planning.
Safe Learning-based Control
We are also exploring methods that combine probabilistically-safe NMPC with policies from reinforcement learning. Our recent research suggests that we can improve the safety of reinforcement learning policies, especially in the case of stochastic dynamics model mismatch.
Multi-Domain and Multimodal Autonomy
Many autonomous robotic systems can be effectively modeled as hybrid dynamical systems, consisting of both continuous and discrete states. This occurs when robotic systems exhibit different autonomous modes, operate across distinct physical domains, or collaborate with one another through discrete actions. Oftentimes, autonomous decision-making decouples discrete and continuous behaviors, but to enable agile robots, we must reason about continuous and discrete behaviors simultaneously.
Multi-Domain Robotics
Robotic systems capable of dynamically transitioning between different operational domains present unique challenges to control and estimation. This is especially true when navigating between different fluid domains. In our prior research, we developed a multi-domain fixed-wing aerial-aquatic vehicle capable of transitioning from water-to-air. Our vehicle leveraged onboard sensing and compute to enable an autonomous water-exit and transition into steady-level flight.
Fixed-Wing Aerial-Aquatic Vehicle
Simultaneous Coordination and Motion Planning
Achieving unified coordination and motion planning in complex real-world environments is a fundamental challenge for multi-robot systems. Oftentimes, these domains are explored in isolation, without the proper attention given to the interactions between the two. In our recent research, we have investigated optimization-based approaches for achieving multi-robot planning and coordination on graphs. Our approach, which formulates an efficient mixed-integer program, can solve for sophisticated coordinated behaviors fast enough to enable real-time replanning for large teams of robots.