Distributed Robotics From Science to Systems

Transcription

1 Distributed Robotics From Science to Systems Nikolaus Correll Distributed Robotics Laboratory, CSAIL, MIT August 8, 2008

2 Distributed Robotic Systems DRS 1 sensor 1 actuator... 1 device Applications Giant, self adapting sensor Working in parallel Providing communication backbone Flexible Structures Advantages Robustness Specialization Speed Overcome specific single robot limitations Future Autonomous sensors and actuators are networked hierarchically Home, City, Environment 2

3 Mobile Sensors and Actuators Mobile sensors and actuators are complex systems Multiple CPUs Multiple communication interfaces Multiple sensors Signal processing, planning and control Varying degrees of Autonomy Sensing Mobility Actuation Collaboration irobot Willow Garage 3

4 Networks of Mobile Sensors and Actuators Internet Foraging Measuring Mowing Plowing Inspecting Cleaning Actuation Actuation Actuation Actuation Environment 4

5 Challenges Inter robot state estimation less reliable than intra robot state estimation Abstract human input > individual robot control Centralized planning > distributed control (robustness, scalability) Distributed sensors > unified view (fusion) Adaptation for improving functionality and robustness 5

6 Coordination Schemes for Multi Robot Teams

7 Possible Coordination Schemes for Multi Robot Systems Best Worst Degree of Coordination (Kalra, PhD 2006) 10/12/2007 Nikolaus Correll 7

8 Ants: Loose Coordination, No Planning Task: find shortest path No plan, purely reactive Local communication 10/12/2007 Direct Indirect (Pheromones) Coordination Jean Louis Deneubourg, ULB Although the ants approach has advantages, they would most likely use a better algorithm if they could (bigger brain, long range communication) Nikolaus Correll 8

9 Yachting: Fully planned, tightly coordinated Coordination 10/12/2007 9

10 Yachting: Fully planned, tightly coordinated Coordination Weather Strategist Grinders Sails Tactician Competition Runner Trim Navigator Helmsman Trimmers Landmarks/Position 10

11 Research Questions How to design the best system for a set of resources and constraints How to predict the performance of a specific system? How do additional capabilities and different reliability change a system s performance? Nikolaus Correll 11

12 Distributed Robotic Systems at DRL/CSAIL

13 Goal: AquaNet (Detweiler, Vasilescu) Long term underwater sensing and monitoring Localization and Tracking Autonomous deployment and recovery Technical challenges: Networking hardware and protocols Autonomous navigation Current Status: Acoustic and optic modems Prototype 10 static and 2 mobile nodes Deployment in French Polynesia

14 Animal Modeling and Control with Virtual Fences (Schwager, Detweiler, Vasilescu, Doniec) Goals: model and control the motion of livestock autonomous Range Management Technical challenges: Collect behavioral and trajectory data Build a data driven model of herding behavior Develop virtual fence control and plan Current Status: Manufactured custom data collection devices Preliminary modeling techniques: classification and parameter identification on cow herds data Demonstrated autonomous gathering Collaboration with Dean Anderson, USDA ARS

15 Goals: Decentralized Swarm Control (Julian) Embedded hardware for swarm intelligence algorithms Multi Camera sensor fusion Decentralized control among multiple fourrotor flying robots Technical challenges: Stable control with limited and/or unreliable environmental sensing Limited onboard resources for embedded hardware Status: Achieved stable waypoint control with Vicon MotionCap system Fabricated prototype control hardware

16 Goal: Intelligent Transportation (Lim) Estimate traffic flow using GPS, cell phone positioning, toll booth transit time Find fast and reliable path using both real time and historical traffic information N=500 Technical Challenges: Noisy positioning data analysis Efficient optimal path finding algorithm Current Status: Establishing road speed DB in Boston N=15 Implemented path search using GPS DB Prototype of 2 nd generation node with OBD2/CAN access (802.11b, GPS, accelerometer) Collaboration with Sam Madden and Hari Balakrishnan, CSAIL

17 Goal: River Flood Prediction on Sensor Predict river flooding 24 hours in advance on autonomous sensor network Technical challenges: Model driven processing in sensor network Data intensive processing at sensor node (data mining, inferencing, and learning) Current Status: Sensing, Computation, and Office nodes developed Field experiments in Honduras and Dover, Massachusetts Network (Basha)

18 Autonomous Deployment and Maintenance of low cost Mesh Networks Goal: Mobile b outdoor mesh networks for < $100/node Technical Challenges: Resource trade offs Miniaturization Routing in dynamic environments Current Status: b routers on irobot Create, Accelerometer, GPS Mesh Networking Deployment algorithm with provable lower bounds 18

19 Jet turbine inspection by a swarm of miniature robots Goal: Autonomous inspection of mechanical structures Technical Challenges: Multiple robots needed for effective inspection Miniaturization Limited design choices Current Status: Extensive studies of coordination schemes using 20 miniature robots Resource trade offs (radio, localization) studied experimentally and analytically in 2D environment 05/28/2008 Willow Garage 19

20 Conclusion and Future Directions Distributed Robotics has made the step from the laboratory into the real world Robots are part of large scale distributed intelligent systems What is needed Make Robotic Prototyping Easy Make Robots Real Cheap Make Robotics More Pervasive in Everyday Life 20