Cooperative Distributed Systems Laboratory Projects

This project is intended to comprehensively develop, test and deliver 6 UAVs with sensors, loggers, communications, and related equipment, operating protocols, and related robot control software packages.

• It to be readily deployed statewide by an ODOT team to facilitate data collection with its payload of sensors
• The project aims at the following:
  
  • Processing of data collected by the UAVs for actionable information by the user for various purposes at Ohio Department of Transportation.
  • Based on data interpretation and UAVs platforms, coordinate with various other teams including but not limited to first responders viz. police, fire, medical, etc.
  • Providing an administrative platform (e.g., via user-friendly but secure website) in a timely manner to control the UAVs in action.

A major issue in the use of cameras mounted on UAVs for fire detection is the presence of smoke that occludes the hot spots in videos. This project focused on reconstructing images from video of scenes occluded by thick smoke and developed a method for filtering out the smoke occlusions in fire image streams using Proper Orthogonal Decomposition (POD). The method works by extracting frames from the video. It is assumed that the images are taken from camera with static background and moving smoke. Using POD, the modes corresponding to the dynamic part of the image sequence representing the smoke are identified and then filtered out to obtain a clear background. The smoke removal technique thus restores the original background that contains the fire information. The technique is applied to a number of sample videos and it is demonstrated that the smoke is sufficiently removed from the video with the background information intact. Following figure shows a sample of original and filtered image obtained from a laboratory experiment

Wildfire growth is typically modeled using a Partial Differential Equation (PDE) that can be solved using grid-based finite difference or finite element methods using some boundary and initial conditions. Given the geographical spread of the environment over which such events occur and with the temporal aspect factored in, data from sensing systems for such processes are very high dimensional in nature, and thus developing a filtering mechanism that incorporates sensing data for online and real-time estimation and prediction for applications can be very challenging due to the computational complexity involved. In this project, we developed a Kalman filtering based method harnessing concepts in Proper Orthogonal Decomposition (POD) for dimensional reduction. The rationale behind the proposed approach is that methods such as the POD facilitate reduced-order modeling, thus reducing the computational complexity. Implemented in a grid-based model of the environment, the approach can eventually be integrated into Geographic Information Systems (GIS) for effective utilization of information and decision making for incident managers. Following figure shows some results and demonstrates the effectiveness in removing noise in measurements.

Control of a large number of distributed systems provides several technical challenges which includes scalability of control and information processing algorithms, stability, and robustness to not only sensor/actuator noises but also to broken/delayed communication. The research conducted in CDS Lab is at the intersection of a number of multidisciplinary aspects (as shown in the figure) including:

1. Systems and controls: Development of decentralized control algorithms, optimization methods, and novel mathematical tools to analyze the stability of system and its robustness to noise and uncertainties.
2. Probability and Statistics: Development of tools to represent uncertainties and for their analysis, specifically in a networked environment.
3. Self-organizing behaviors: Drawing inspirations from systems seen in nature at microscopic levels such cellular behaviors to that seen at macroscopic levels such flock of birds, swarm of ants, and schools of fish.

Researchers at the CDS lab have been extensively involved in various aspects of research on multi-robot cooperative control including work on aggregation/segregation of heterogeneous units in robotic swarms, development of control laws for a robotic swarm emulating ant foraging behavior, noise induced adaptive emergent behaviors in swarm robotic systems, and performance driven decentralized control of robotic agents. Some of current focus of research in CDS Lab include: study of statistical-mechanical concepts such random graphs, random geometric graphs, and mean-field theory to model and analyze system behaviors; and study of the effect of noise on robust self-organization of robotic swarms.

It is often difficult to obtain precise information about the states of large-scale systems due to non-linearities involved, complex interactions, and uncertainties. It may, however, be possible to analytically obtain average quantities that would provide crude representations of complex behaviors. For example, in a study on segregation of heterogeneous units in a swarm of robotics agents, average distances between agents of similar and dis-similar types were used to obtain analytical results on segregation of heterogeneous agents under very simple control laws based on differential potential.

A significant challenge in swarm robotics is the design and control of a robotic swarm capable of adaptive behavior dictated by local communication in uncertain environments. Given their remarkable propensity to adapt to rapidly changing environments, the dynamics of biological systems, such as an ant colony system, provide fundamental insights in this context. This research focused on development of a new mathematical model, represented by coupled Partial Differential Equations inspired by Keller Segel model of bacterial chemotaxis, for ant foraging that accounts for different behaviors exhibited by foragers in search of food and food carrying ants. The model essentially shows the evolution of i) food searching (foraging) ants; ii) food carrying ants; and iii) the pheromone distribution in the space. Food search is governed by an environmental potential, the pheromone gradient and is also characterized by inherent randomness that allows for a comprehensive search of the entire domain for food sources. Moreover, a fraction of the foraging ants change character to food carrying ants within a certain neighborhood of the food sources and the reverse happens within a neighborhood of the nest.

This project focused on developing cooperative control laws for cyclic pursuit of robotic agents to track a closed perimeter. In this work, a linear interaction law is proposed for pursuit with UAV. Using principles of linear control theory, stability conditions were obtained and it was shown that, within specified stability regions, the system was robust to addition or deletion of agents. Similarly, cooperative control algorithms for wildfire monitoring and fighting using non-linear interaction potential is also developed.

On a research studying the effect of randomness on robust flocking behaviors in multi-robotic agent systems, we were able to show that single-cluster flock would asymptotically form if designed randomness is introduced in the system. It was concluded via analysis as well as extensive simulations that randomness provided a necessary mechanism for robust flocking behavior.

The objective of this project is to design and build a fully autonomous quad-rotor called Smart Unmanned Aerial for Exploration (SUAVE) which will be used to participate in the Autonomous Aerial Vehicle Competition at the 2014 Ohio UAS Conference to be held in Dayton, Ohio from Aug. 26-28, 2014. The project will utilize 3D Robotics airframe and off-the-shelf sensors, wireless communication devices, motors, micro-controllers, and some specialized sensors such as laser scanner. The project will involve integrating all the components, designing fixtures for specialized components, designing software interfaces for different sensor and communication devices with the onboard and off-board computers, developing localization and navigation algorithms for indoor environments, and finally extensively testing the system in the laboratory environment.

Complex networked systems represent several key engineering infrastructures in our modern society including power grids, wireless networks, communication and transportation networks, world-wide web, and cloud computing networks. In this context, of particular interest is the determination of optimal operation point of these networks to ensure their efficiency/performance, robustness to uncertainties, and responsiveness to dynamic changes. There are several unique challenges to solving the optimization problems in these large-scale networks, such as high-dimensionality due to the NP-Hard nature of many of such problems and lack of global information which traditional centralized optimization methods fails to handle. To address the challenges of high-dimensionality and lack of global information, recent years have seen the emergence of the concept of decentralized or distributed optimization. However, solving the global optimization problem in a distributed and asynchronous fashion using local, and often incomplete and noisy, information in presence of local and global constraints is still an open problem in literature.

This research direction concerns development and application of novel distributed optimization methods for large-scale networked systems. One of the main focuses of this research is to develop a theoretical foundation for carrying out optimization in a distributed framework using only local information specifically focusing on issues of high-dimensionality, non-convexity, slow convergence, and adaptability to dynamic changes for the large-scale networked systems.

Previous research in this area included development of Newton based primal-dual interior point distributed optimization methods for Network Utility Maximization problem in the networks, and application of Market-Based methods for optimal power flow in power grids and resource allocation in cloud computing systems.

In this work the features of the robot environment such as lane markers and obstacles are divided in to two classes using k-NN algorithm and a maximum margin hyperplane is obtained using SVM that optimally divides both the classes. This hyperplane represents a collision free path. In situations that demand both lane compliance and obstacle avoidance, the use of this technique would simplify the problem and eliminate any conflicts that may arise due to the usage of separate algorithms.

Mobile robot navigation in hilly and uneven terrain is a challenging problem. However, humans possess an uncanny knack to identify paths even in difficult terrains. Therefore, it is beneficial to use this human expert knowledge in identifying paths from start to goal. These paths can be considered as additional information to the robot to help them move towards the goal. Objective of this research was to develop a framework where the 1) Robot can make use of human expert knowledge to easily reach the goal from any start position; and 2) The robot can autonomously navigate through the terrains even in the absence of expert assistance. This research made use of reinforcement learning techniques to learn from human experts who provided rough trajectories as inputs. Fuzzy logic was further used to obtain detailed trajectory which was safe for the robot.

Development of an exoskeleton for sit-to-stand (STS) transition support based on multimodal action intent recognition

1.5 million Senior citizens live under supervision and most require assistance with at least one or more Activities of Daily Living (ADL), including transferring in and out of chairs, beds and toilets which requires the ability to perform sit-to-stand transition. This sit-to-stand transition is a complex full-body activity that requires the synergistic coordination of the upper and lower limbs and trunk. The goal of this research is to come up with a working prototype of an active, assistive exoskeleton which can be controlled based on behavioral models of user’s intent. The research plan includes synchronized multimodal data-collection of sit-to-stand transitions across various environmental situations and action intent contexts and development of intelligent control algorithms to actuate and operate the exoskeleton. This work can be expanded to control robots in any environment which requires human-robot coordination to complete the same task.