@ MapleCG
papers
_____
cfp
= {member}
Assistant Professor > CS > NUS
Ph.D. > ECE ◊ RI > CMU
MapleCG > CS > NUS
lowkhATcompDOTnusDOTeduDOTsg
|
research @ MapleCG |
|
_____ papers _____ |
|
--> cfp | |
| aug 1, 2011 | | |
|
|
|
[ homee ] | |
* foto * | |
teach +++ | | |
|
MapleCG = {member} |
|
| cv || | |
LOW, BRYAN KIAN HSIANG |劉謙雄| Assistant Professor > CS > NUS Ph.D. > ECE ◊ RI > CMU MapleCG > CS > NUS lowkhATcompDOTnusDOTeduDOTsg | |||
|
@ |
research MapleCG |
Multi-Agent Planning, Learning, and Coordination Group (MapleCG) _____| c u r r e n t |____________ |
|
|
ACTIVE MARKOV-BASED ROBOTIC EXPLORATION & MAPPING FOR ENVIRONMENTAL SENSING PROJECT DURATION : Jan 2010 - Present PROJECT AFFILIATION : Collaborative Multi-Robot Exploration of the Coastal Ocean (Collaborators: John M. Dolan, CMU; Gaurav S. Sukhatme, USC; Kanna Rajan, MBARI) PROBLEM MOTIVATION Research in multi-robot exploration and mapping has recently progressed from building occupancy grids to sampling spatially varying environmental fields (e.g., plankton density, pollutant concentration, temperature fields) that are characterized by continuous-valued, spatially correlated measurements. Exploration strategies for building occupancy grid maps usually operate under the assumptions of (a) discrete, (b) independent cell occupancies, which impose, respectively, the following limitations for learning environmental field maps: these strategies (a) cannot be fully informed by the continuous field measurements and (b) cannot exploit the spatial correlation structure of an environmental field for selecting observation paths. As a result, occupancy grid mapping strategies are not capable of selecting the most informative observation paths for learning an environmental field map. Furthermore, occupancy grid mapping strategies typically assume that range sensing is available. In contrast, many in situ environmental sensing applications permit only point-based sensing, thus making a high-resolution sampling of the entire field impractical in terms of resource costs (e.g., energy consumption, mission time). In practice, the resource cost constraints restrict the spatial coverage of the observation paths. Fortunately, the spatial correlation structure of an environmental field enables a map of the field (in particular, its unobserved areas) to be learned using the point-based observations taken along the resource-constrained paths. To learn this map, the field is represented using a rich class of Bayesian non-parametric models called the Gaussian process (GP). More importantly, the GP model allows an environmental field to be formally characterized and consequently provides formal measures of mapping uncertainty (e.g., based on mean-squared error [AAMAS-08] or entropy [ICAPS-09]) for directing a robot team to explore highly uncertain areas of the field. How then does a robot team plan the most informative resource-constrained observation paths to minimize the mapping uncertainty of an environmental field? To address this, our ICAPS-09 work has proposed an information-theoretic multi-robot exploration strategy that selects non-myopic observation paths with maximum entropy. When this strategy is applied to sampling a GP-based field, it can be reduced to solving a non-Markovian, deterministic planning problem called the information-theoretic multi-robot adaptive sampling problem (iMASP). Due to the non-Markovian problem structure of iMASP, its state size grows exponentially with the length of planning horizon. To alleviate this computational difficulty, it can be solved approximately using an anytime heuristic search algorithm called Learning Real-Time A* or a greedy algorithm (i.e., by solving the myopic formulation of iMASP). However, they inherit iMASP's non-Markovian problem structure and consequently scale poorly with the length of history of observations. Hence, it becomes computationally impractical to use these non-Markovian path planning algorithms for in situ, real-time active sampling. PROPOSED METHODOLOGY To ease this computational burden, this work proposes a principled Markov-based approach to efficient information-theoretic path planning for active sampling of GP-based environmental fields, which we develop by assuming the Markov property in iMASP planning. To the probabilistic robotics community, such a move to achieve time efficiency is probably anticipated. However, the Markov property is often imposed without carefully considering or formally analyzing its consequence on the performance degradation while operating in non-Markovian environments. In particular, to what extent does the environmental field structure affect the performance degradation due to violation of the Markov assumption? Motivated by this lack of treatment, this work is novel in demonstrating both theoretically and empirically the extent of which the degradation of active sampling performance depends on the spatial correlation structure of an environmental field. An important practical consequence is that of establishing environmental field conditions under which the Markov-based approach performs well relative to the non-Markovian iMASP-based policy while enjoying significant computational gain over it. PUBLICATIONS
|
|
 TAOSF PROSPECT |
MULTI-ROBOT ADAPTIVE SAMPLING FOR ENVIRONMENTAL SENSING & MONITORING PROJECT DURATION : Jul 2005 - Mar 2010 PROJECT LAB : T-SAR PROBLEM MOTIVATION Recent research in multi-robot exploration and mapping has focused on sampling environmental fields, some of which typically feature a few small hotspots in a large region. Such a hotspot field often arises in two real-world applications: (1) planetary exploration such as geologic reconnaissance and prospecting for mineral deposits or natural gases, and (2) environment and ecological sensing such as precision agriculture, and monitoring of ocean phenomena (e.g., plankton bloom, anoxic zones), forest ecosystems, rare species, pollution (e.g., oil spill), or contamination (e.g., radiation leak). In particular, the hotspot field is characterized by continuous-valued, spatially correlated measurements with the hotspots exhibiting extreme measurements and much higher spatial variability than the rest of the field. With limited (e.g., point-based) robot sensing range, a complete coverage becomes impractical in terms of resource costs (e.g., resource consumption). So, to accurately map the field, the hotspots have to be sampled at a higher resolution. The hotspot field discourages static sensor placement because a large number of sensors has to be positioned to detect and refine the sampling of hotspots. If these static sensors are not placed in any hotspot initially, they cannot reposition by themselves to locate one. In contrast, a robot team is capable of performing high-resolution hotspot sampling due to its mobility. Hence, it is desirable to build a mobile robot team that can actively explore to map a hotspot field. PROPOSED METHODOLOGY To learn a hotspot field map, the exploration strategy of the robot team has to plan the most informative resource-constrained observation paths that minimize the uncertainty of mapping the hotspot field. By representing the hotspot field using rich classes of Bayesian non-parametric models such as the Gaussian process or log-Gaussian process, formal measures of mapping uncertainty (e.g., based on mean-squared error [AAMAS-08] or entropy [ICAPS-09] criterion) can be defined and subsequently exploited by our proposed adaptive sampling algorithms for directing the robot team to explore highly uncertain areas of the field. In contrast to non-adaptive sampling strategies that only perform well with smoothly-varying fields, our non-myopic adaptive sampling algorithms can exploit clustering phenomena (i.e., hotspots) to plan observation paths that produce lower mapping uncertainty. PUBLICATIONS
|
|
|
DISTRIBUTED
LAYERED ARCHITECTURE FOR SELF-ORGANIZING MOBILE SENSOR NETWORKS PROJECT DURATION : Nov 2002 - Jun 2005 PROBLEM MOTIVATION One of the fundamental issues that arises in a sensor network is coverage. Traditionally, network coverage is maximized by determining the optimal placement of static sensors in a centralized manner, which can be related to the class of art gallery problems. However, recent investigations in sensor network mobility reveal that mobile sensors can self-organize to provide better coverage than static placement. Existing applications have only utilized uninformed mobility (i.e., random motion or patrol). In contrast, our work here focuses on informed, intelligent mobility to further improve coverage. Our network coverage problem is motivated by the following constraints that discourage static sensor placement or uninformed mobility: (a) no prior information about the exact target locations, population densities or motion pattern, (b) limited sensory range, and (c) very large area to be observed. All these conditions may cause the sensors to be unable to cover the entire region of interest. Hence, fixed sensor locations or uninformed mobility will not be adequate in general. Rather, the sensors have to move dynamically in response to the motion and distribution of targets and other sensors to maximize coverage. Inspired by robotics, the above problem may be regarded as that of low-level motion control to coordinate the sensors' target tracking movements in the continuous workspace. Alternatively, it can be cast as a high-level task allocation problem by segmenting the workspace into discrete regions such that each region is assigned a group or coalition of sensors to track the targets within. PROPOSED METHODOLOGY This work presents a reactive layered multi-robot architecture for distributed mobile sensor network coverage in complex, dynamic environments. At the lower layer, each robot uses a reactive motion control strategy known as Cooperative Extended Kohonen Maps to coordinate their target tracking within a region without the need of communication. This strategy is also responsible for obstacle avoidance, robot separation to minimize task interference, and navigation between regions via beacons or checkpoints plotted by a motion planner. At the higher layer, the robots use a dynamic ant-based task allocation scheme to cooperatively self-organize their coalitions in a decentralized manner according to the target distributions across the regions. This scheme addresses the following issues, which distinguish it from the other task allocation mechanisms: Task Allocation for Multi-Robot Tasks: Existing algorithms (e.g., auction-and behavior-based) assume a multi-robot task can be partitioned into single-robot tasks. But this may not be always possible or the multi-robot task can be more efficiently performed by coalitions of robots. Coalition Formation for Minimalist Robots: Existing coalition formation schemes require complex planning, explicit negotiation, and precise estimation of coalitional cost. Hence, they do not perform well in dynamic, real-time scenarios. Cooperation of Resource-Limited Robots: Robots with limited communication and sensing capabilities (i.e., partial observability) can only obtain local, uncertain information of the dynamic environment. With limited computational power, their cooperative strategies cannot involve complex planning or negotiations. PUBLICATIONS
Coverage of 30 targets (green) with 15 ant robots (white)
|
|
|
ACTION SELECTION MECHANISM FOR MULTI-ROBOT TASKS PROJECT DURATION : Sep 2002 - Nov 2002 PROBLEM MOTIVATION A central issue in the design of behavior-based control architectures for autonomous mobile robots is the formulation of effective mechanisms to coordinate the behaviors. These mechanisms determine the policy of conflict resolution between behaviors, which involves behavioral cooperation and competition to select the most appropriate action. The actions are selected so as to optimize the achievement of the goals or behavioral objectives. Developing such an action selection methodology is non-trivial due to realistic constraints such as environmental complexity and unpredictability, and resource limitations, which include computational and cognitive capabilities of the robot, incomplete knowledge of the environment, and time constraints. As a result, action selection can never be absolutely optimal. Given these constraints, the action selection scheme should be able to choose actions that are good enough to satisfy multiple concurrent, possibly conflicting, behavioral objectives. PROPOSED METHODOLOGY Our motivation of the action selection mechanism is to develop a motion control strategy for autonomous non-holonomic mobile robots that can perform distributed multi-robot surveillance in unknown, dynamic, complex, and unpredictable environments. By implementing the action selection framework using an assemblage of self-organizing neural networks, it induces the following key features that significantly enhance the agent's action selection capability: self-organization of continuous state and action spaces to provide smooth, efficient and fine motion control, and action selection via the cooperation and competition of Extended Kohonen Maps to achieve more complex motion tasks: (1) negotiation of unforeseen concave and narrowly spaced obstacles, and (2) cooperative tracking of multiple mobile targets by a team of robots. Qualitative and quantitative comparisons for single- and multi-robot tasks show that our framework can provide better action selection than do potential fields method. PUBLICATIONS
|
|
|
INTEGRATED ROBOT PLANNING AND CONTROL PROJECT DURATION : Jul 2001 - Sep 2002 PROBLEM MOTIVATION Robot motion research has proceeded along two separate directions: high-level deliberative planning and low-level reactive control. Deliberative planning uses a world model to generate an optimal sequence of collision-free actions that can achieve a globally specified goal in a complex static environment. However, in a dynamic environment, unforeseen obstacles may obstruct the action sequence, and replanning to react to these situations can be too computationally expensive. On the other hand, reactive control directly couples sensed data to appropriate actions. It allows the robot to respond robustly and timely to unexpected obstacles and environmental changes but may be trapped by them. PROPOSED METHODOLOGY The problem of goal-directed, collision-free motion in a complex, unpredictable environment can be solved by tightly integrating high-level deliberative planning with low-level reactive control. This work presents two such architectures for a nonholonomic mobile robot. To achieve real-time performance, reactive control capabilities have to be fully realized so that the deliberative planner can be simplified. These architectures are enriched with reactive target reaching and obstacle avoidance modules. Their target reaching modules use indirect-mapping Extended Kohonen Map to provide finer and smoother motion control than direct-mapping methods. While one architecture fuses these modules indirectly via command fusion, the other one couples them directly using cooperative Extended Kohonen Maps, enabling the robot to negotiate unforeseen concave obstacles. The planner for both architectures use a slippery cells technique to decompose the free workspace into fewer cells, thus reducing search time. Any two points in the cell can still be traversed by reactive motion. PUBLICATIONS
|
