Ali-akbar Agha-mohammadi

Transcription

1 Ali-akbar Agha-mohammadi Parasol lab, Dept. of Computer Science and Engineering, Texas A&M University Dynamics and Control lab, Dept. of Aerospace Engineering, Texas A&M University Statement of Research Broad research interests and related applications: My research is in the area of stochastic systems, computational methods for sensing, estimation, and control problems, with a focus on robotic systems. In real-world systems, uncertainty (both in the system model and sensory readings) is omnipresent. A few of the vast set of applications for this research include (i) Motion planning for mobile robots and manipulators under uncertainty, (ii) precise control of medical robots under imperfect measurements, (iii) multiagent cooperation and coordination under stochastic communication, and (iv) planning for robots aiding healthcare operating in changing and uncertain environments with people. Approach/philosophy of research & existing challenges: Specifically, I am interested in principled tools that can robustly deal with such problems in a probabilistic setting, such as stochastic control methods, filtering and estimation theory, and AI-based decision making methods under uncertainty. Markov Decision Processes (MDP) and Partially-Observable MDPs (POMDP) are one of the most principled general frameworks that model the problem of sequential decision making under uncertainty. However, the computational intractability of these frameworks limits their usage to problems with small discrete state spaces and restricts their applicability in realistic applications. For example, the difficulty of the problem of motion planning under uncertainty usually leads to solutions where the planning in the information space is ignored or the planning and control phases are separated. Such solutions may result in overly conservative or unreliable behaviors in general. Research goals and contributions: My main research goal is to establish missing links between theory and practice in planning under uncertainty by creating principled tools and algorithms with an emphasis on robustness, reliability, and scalability. In particular, in my PhD dissertation, I investigate the connections between the computational approaches to solve MDP and POMDP problems with their analytic counterpart approaches in control theory (see Fig. 1). I show that establishing such a connection can solve many problems in the domain of planning and control under uncertainty, and more importantly can lead to principled theoretical tools that are applicable to practical systems. Accordingly, my current research is highly interdisciplinary, lying in the intersection of control systems, computational methods and my PhD dissertation is conducted jointly at the Parasol lab (CSE Dept) and Dynamics and Control Lab (Aerospace Dept.). Different tools in dealing with stochastic systems; I believe research in the intersection of these areas can solve important unsolved problems, in particular in the domain of "planning and control under uncertainty".

2 My Dissertation Research Problem: My PhD research focuses on designing planning and control schemes for robotic systems under process noise and imperfect measurements. This problem calls for control and planning in the space of probability distributions (since the exact state of the system is unknown), which is referred to as "belief space" or "information space." One of the most general formulations of this problem is the POMDP formulation. Planning motions (feedback law) for a robotic system to reach a desired goal point (in the presence of state and control constraints) under uncertainty is a well-known instance of this problem. Challenges: However, POMDPs are notorious for their computational intractability, and the reported solutions are usually limited to problems with small discrete state spaces. For example, a problem with a 1000 discrete states has a 1000 dimensional belief space! Furthermore, a problem with a continuous state space (where almost all real-world robotic problems reside) has an infinite dimensional belief space. State-of-the-art: Inspired by the success of sampling-based algorithms in deterministic settings, different methods have attempted to sample points (graph nodes) in the belief space and approximate the belief space using a graph structure to reduce the complexity of this space. However, constructing such a graph in belief space is challenge due to the difficult task of sampling graph nodes (characterizing reachable beliefs) and connecting them to each other. Therefore, the results in literature are limited to trees that are only valid for a single initial belief, and are not robust to model errors and large deviations. Also the computational complexity of constructing such trees in belief space is exponential in the number of underlying samples, which restricts their usage in large problems. Contributions of my PhD research: In my dissertation, establishing a connection between stochastic feedback controllers, AI-based POMDP approaches, and randomized algorithms led us to design the first/only "roadmap" (graph) in the belief space [1,6] that has all the benefits of its celebrated counterpart, i.e., PRM (Probabilistic Roadmap Method) in the configuration/state space. Basically, using locally distributed feedback controllers, such as Linear Quadratic Gaussian (LQG) controllers, and designing a novel switching method among them, we characterize the space of reachable beliefs. By sampling this space, we can generate a graph in the belief space. Accordingly, the POMDP problem is reduced to a tractable optimal control problem on this graph. This construct exhibits desirable properties sought in the robotics (motion planning) community, such as robustness, reliability, scalability, and real-time (feedback) plan execution. We believe that this research will have a great impact on the robotics society as it fills the gap between the POMDP theory and practice. Theoretical highlights: The Feedback-based Information RoadMap (FIRM), which is a graph in belief space whose nodes are certain probability distributions (belief) and whose edges are certain feedback controllers that guarantee belief reachability [1,6]. Robust (to large deviations), reliable (constraints can seamlessly incorporated), and scalable planning schemes for stochastic systems. Characterization of the reachable belief manifold using different controllers such as Linear Quadratic Gaussian (LQG) controllers [1,6], Dynamic Feedback Linearization-based (DFL) controllers [4], or Periodic LQG controllers (for kinodynamical planning) [7]. Theoretical guarantees such as success probability or probabilistic completeness on the obtained solution [5]. A unified graph-based control scheme in the absence and presence of measurement noise [2,3]. Practical highlights: Achieving desirable properties, such as robustness, reliability, and scalability, we were able to conduct the first experiments on implementing robust belief space planners (i.e., with real-

3 time replanning capabilities) on physical robots (i-robot creates) under motion noise and noisy sensory readings [9]. A unique feature of our implementation is its robustness to the changing environment, sensor failures, and large deviations. The main goal of the experiments is to minimize collision probability while maximizing the information gain obtained from sensors (camera in our tests) and show the robustness of the systems to different types of changes and failures. We believe this is a significant step in making POMDP methods a practical tool for robotic applications under uncertainty. Future research I would like to explore a broad range of topics within the area of stochastic systems, planning and control under uncertainty. In particular, I like to see closure between theory and physical realization. My PhD research paves the way to tackle many challenging problems in planning under uncertainty for real-world robotic systems. It also provides many opportunities for future collaborations and new projects. Some specific future research directions that I am planning to pursue are outlined in the following paragraphs. In most of the following projects, I have already taken some initial steps. 1) Planning in dynamic uncertain environments: Although the SLAM (Simultaneous Localization And Mapping) problem is a well-studied problem in AI and Robotics, the SPLAM (Simultaneous Planning, Localization And Mapping) and the question of how a robot has to be moved and controlled to gain maximum possible information from the unknown environment has not been fully answered yet. Moreover, the environment may be dynamic and change over time. Typical applications where this problem arises are service robotics, home healthcare aids for the elderly, or search and rescue robotics. Another important application is in the control of mobile manipulators (e.g., PR2 or Kuka YouBot), where imperfect sensory readings and partial knowledge about the environment introduce many interesting problems. Using feedback controllers and exploiting the FIRM framework in a SLAM setting, we can tackle the SPLAM problem. For dynamic environments, I believe the information roadmaps proposed in my dissertation are a very promising framework to handle this problem since the edges of an information roadmap are independent of each other, and hence, local changes (dynamicity) in different parts of the environment only require local updates of the information graph, which is a feasible operation in real time. We have taken initial steps in this direction. Very recently, we have implemented the first information space planner on physical robots operating in a changing office environment [9]. Furthermore, my experience with Rescue Robotics, SLAM, and physical robots [11-16] equips me with more tools to tackle the SPLAM problem. 2) Multi-robot systems and cooperative planning under uncertainty: Another main area I would like to work on is planning for multi-agent systems under uncertainty. Taking uncertainty into account in communication between robots and in cooperative planning (e.g., cooperative load transport) leads to behaviors that are beyond the abilities of individual robots, and also robust to failures. There are many unsolved problems in this domain ranging from the modeling and quantification of the uncertainty to designing cooperation protocols and decentralized plan execution. In this direction, I have worked on multi-agent systems and have started to design an information space planner for multiagent systems. We have conducted some initial simulations on a team of robots with heterogeneous sensing abilities. The initial results are promising and have led to interesting team behaviors, which have significantly increased the success probability of the mission compared to the success probability obtained by the team ignoring the stochasticity of the communication among team members.

4 3) Planning in non-gaussian information spaces: In many robotic applications, the probability distribution of the system state may be multi-modal and non- Gaussian. For example, when the data association (i.e., matching predicted features and measure features) is vague, the position of the robot or objects in the environment may be described by a multi-modal distribution. Planning in such information spaces is an open problem. Generating information roadmaps based on particle filters is an interesting future research direction. Also, it is well-known that the evolution of the state pdf (probability distribution function) in a nonlinear system is governed by the Fokker- Planck-Kolmogorov's (FPK) partial differential equation. Thus, it is an interesting direction to quantify the uncertainty and make information roadmaps based on FPK. I have taken some initial steps in this direction. An important application of such a framework can be in planning based on vision sensors or laser range finders, where we extract natural features from the environment and the data association in this setting may lead to multi-modal distributions. My experience with robot vision, cameras, and laser range finders [10-16] will significantly help in following this future direction. 4) Medical Robotics: Due to their sensitivity, robot-assisted medical applications need more reliable and accurate control strategies. Thus, methods that take into account uncertainty are very desirable in medical applications as they can increase accuracy and safety. I have taken some initial steps in robot-assisted medical needle steering [8] as a testbed for the information roadmap methods proposed in my thesis. Magnetically Guided Capsule Endoscopy is another interesting testbed for stochastic control methods. It is used to examine parts of the gastrointestinal tract that cannot be seen with other types of endoscopy. Finally, robotic prosthesis is a fascinating direction in which stochastic control can play an important role. It has been shown that adding a level of machine autonomy can significantly improve the performance of a Brain-Machine Interface (BMI) system. As it is aligned very well with my background, I am interested in exploring how to add such a level of autonomy to increase the safety, efficiency, ease of use, robustness, and adaptability of the control and sensing algorithms used in brain-controlled robotic prostheses. References [1] A. Agha-mohammadi, Suman Chakravorty, Nancy Amato, "Sampling-based Feedback Motion Planning Under Motion Uncertainty and Imperfect Measurements", International Journal of Robotics Research (IJRR), accepted. [2] A. Agha-mohammadi, Sandip Kumar, Suman Chakravorty, "Motion Planning under Uncertainty", A book chapter in Intelligent Systems, AIAA Progress in Aero & Astro Series, Editor: John Valasek, In print. [3] A. Agha-mohammadi, Suman Chakravorty, Nancy Amato, "Graph-based Stochastic Control with Constraints: A Unified Approach with Perfect and Imperfect Measurements", in Proc. American Control Conference (ACC), invited session on Stochastic Models, Control and Algorithms in Robotics, Washington, DC, June [4] A. Agha-mohammadi, Suman Chakravorty, Nancy M. Amato, "Sampling-based Nonholonomic Motion Planning in Belief Space via Dynamic Feedback Linearization-based FIRM", in Proc. IEEE Int. Conf. Intel. Rob. Syst. (IROS), Vilamoura, Portugal, Oct (44% acceptance rate) [5] A. Agha-mohammadi, Suman Chakravorty, Nancy M. Amato, "On the Probabilistic Completeness of the Sampling-based Feedback Motion Planners in Belief Space," in Proc. IEEE Int. Conf. Robot. Autom. (ICRA), Saint Paul, Minnesota, May (40% acceptance rate) [6] A. Agha-mohammadi, Suman Chakravorty, Nancy M. Amato, "FIRM: Feedback Controller-Based Information- State Roadmap, A Framework for Motion Planning Under Uncertainty," in Proc. IEEE Int. Conf. Intel. Rob. Syst.

5 (IROS), pp , San Francisco, CA, Sep (32% acceptance rate) [7] A. Agha-mohammadi, Suman Chakravorty, Nancy M. Amato, "Online Replanning in Belief Space for Dynamical Systems: Towards Handling Discrete Changes of Goal Location, A paper/oral-presentation in IEEE ICRA 2013 Workshop on Combining Task and Motion Planning. [8] A. Agha-mohammadi, Suman Chakravorty, Nancy M. Amato, "Medical Needle Steering under Motion and Sensor Noise using Feedback-based Information Roadmaps," A poster-presentation in IEEE ICRA 2012 Needle Steering Workshop, Saint Paul, Minnesota, May [9] A. Agha-mohammadi, et al., "Dynamic Replanning in Belief Space: An Experimental Study on Physical Mobile Robots", in preparation. [10] A. Agha-mohammadi, Dezhen Song, "Robust Recognition of Planar Mirrored Walls Using a Single View", in Proc. IEEE International Conference on Robotics and Automation (ICRA 11), pp , Shanghai, China, May (49% acceptance rate) [11] A. Tamjidi, Hamid D. Taghirad, A. Agha-mohammadi, "On the Consistency of EKF-SLAM: Focusing on the Observation Models", in Proc. IEEE Int. Conf. Intel. Rob. Syst. (IROS), pp , St. Louis, US, Oct (54.5% acceptance rate) [12] A. Agha-mohammadi, A. Tamjidi, Hamid D. Taghirad, "A Solution for SLAM through Augmenting Vision and Range Information", in Proc. IEEE Int. Conf. Intel. Rob. Syst. (IROS), pp , Nice, France, Oct (48% acceptance rate) [13] A. Agha-mohammadi, A. Tamjidi, Hamid D. Taghirad, "SLAM Based On the LRF Information as the Only Data Source", in Proc. 17 th International Federation of Automatic Control, (IFAC 08), pp , Seoul, Korea, July [14] A. Agha-mohammadi, Hamid D. Taghirad, A. Tamjidi, and Ehsan Mihankhah, "Feature-Based Range Scan Matching For Accurate and High Speed Mobile Robot Localization", in Proc. Third European Conference on Mobile Robots (ECMR 07), Freiburg, Germany, pp , [15] "Multisensor Fusion methods for Solving the SLAM Problem in the EKF Framework", M.Sc. Thesis, Department of Electrical and Computer Engineering, Khaje Nasir Toosi University of Technology, July 2008 [16] Arash Kalantari, Ehsan Mihankhah, A. Agha-mohammadi, "Resquake, A Tracked Mobile Rescue Robot, Rescue Robotics Camp, An oral presentation in IEEE International Workshop on Safety Security and Rescue Robotics (SSRR 07), Rome, Italy, 2007.