Multi-Agent Mapping & Inspection
We are developing a GPS-denied, multi-agent mapping and inspection capability, funded by Lockheed Martin. We employ in-house-developed quadrotor aerial vehicles and small ground rovers equipped with depth-sensing cameras to collaboratively perform simultaneous localization and mapping (SLAM) on a target object, as well as detect features of interest on the target.
Seeker is a NASA Johnson Space Center (JSC) program that aims to demonstrate the core capabilities required for safe external robotic free-flyer inspection of crewed space vehicles. The program consists of a series of cubesat missions that will incrementally add capability. Dr. Akella is serving as the Principle Investigator for the vision system for Seeker. The Seeker-1 technology demonstration mission launched on 17 April, 2019 onboard the Northrop Grumman Cygnus Enhanced resupply vehicle to the International Space Station (ISS). Seeker-1 deployed from Cygnus after Cygnus completed its resupply mission on 16 September, 2019.
In collaboration with the Texas Spacecraft Lab (TSL), we are developing a visual navigation system (VizNav) for Seeker based on solely monocular camera input. VizNav identifies whether or not the target vehicle (Cygnus) is in frame, and if so, determine the relative pose (translation and attitude) of Cygnus with respect to Seeker. A machine learning approach via Convolutional Neural Networks (CNNs) is taken for object detection and recognition, while classical estimation techniques are employed for relative pose estimation.
Hypersonic Trajectory Generation using Indirect Optimal Control Methods
The goal of this research is to rapidly generate high-quality trajectories for hypersonic glide vehicles using indirect optimal control methods. Hypersonic trajectories are generally difficult to solve using indirect methods due to their numerical stiffness. However, continuation and homotopy methods are generally able to generate high-quality trajectories by transforming an a priori known solution into a desired solution. We are developing a computationally efficient application of the stabilized continuation boundary-value-problem solver for hypersonic applications.
Advanced Teaming Architectures for Swarms of Heterogeneous Robots
A major innovation sought through this research is the development of decentralized self-organization algorithms among multiple heterogeneous robotic assets deployed inside contested environments that are robust to emerging threats, lost links, or change in mission priorities. The key feature of our advanced teaming research is the incorporation of spatial representations for task and domain decomposition while accommodating for high-level mission goals such as time-to-contact and topological connectivity (coverage). Scalable approaches for perception will also be implemented to allow for rapid evaluation and recognition of previously detected landmarks that permit real-time answers to: “Have I been here before?” Extended consensus bundle algorithms for bidding and consensus will be formulated such that loosely connected vehicles can self-localize and navigate about landmarks, seamlessly accommodate for merging and splitting of a subset of vehicles, evade obstacles and a priori unknown threats in a scalable fashion while keeping pace with tactical operations.
SERPENT (Satellite Evaluation of Relative Pose Estimation of a Noncooperative Target) is a mission selected by the Air Force Research Laboratory’s University Nanosatellite Program (UNP). Dr. Akella serves as the Principal Investigator of SERPENT, which is being developed in conjunction with the Texas Spacecraft Lab (TSL). The mission proposes to develop an autonomous pose estimation and navigation algorithm by employing convolutional neural networks (CNNs), and computer vision techniques.
Micro Aerial Vehicle (MAV) Guidance
We develop and implement guidance algorithms for micro aerial vehicles (MAVs). For example, we have developed a real-time minimum-snap algorithm for quadrotors using feed-forward neural networks (NNs). The optimal minimum-snap algorithm involves an expensive gradient descent calculation and therefore cannot be performed in real-time. Other techniques such as the trapezoidal velocity profile methods are far from optimal. NNs trained using the output from optimal gradient descent methods result in a significant improvement in the optimality of the trajectories found while also allowing them to be generated in real-time.