Title: Cooperative Algorithms for Mobile Robots and a Sensor Network
1Cooperative Algorithms for Mobile Robots and a
Sensor Network
?
by Maxim Batalin
?
T2
Supervisor (USC) Gaurav Sukhatme
2Contents
- Introduction
- Problems and solutions with Mobile Robots (MR)
and Sensor Networks (SN) - Coverage and Exploration through Sensor Network
Deployment - Sensor Network Repair and Maintenance Implicit
- Probabilistic Navigation
- Task Allocation Problem (scheduling)
- DINTA
- DINTA-MF
- Sensor Network Repair and Maintenance - Explicit
- Current Work Task Allocation for NIMS
- Summary
3Introduction
4Why use both Mobile Robots and Sensor Networks?
- Sensor Network provides distributed
- Sensing
- Computation
- Communication
- Ubiquitous computing is an active area of
research and investment, hence can utilize
intelligence which will be present in an
environment in the future - At the same time Mobile robots deploy, repair and
maintain the network, while accomplishing tasks
that require mobility
Robots can be simpler, cheaper and dont need as
many Can solve wider variety of problems
5Overall objectives
- We are motivated by the idea that MRs and SN each
could leverage strengths from the other - Goal Build a collaborative symbiotic system in
which mobile robots and a sensor network solve
tasks cooperatively and coexist benefiting each
other - Assumptions
- Global information is not accessible (no GPS, no
map, etc.) - Neither robot localization nor mapping is
performed hence robots can be simple - Environment can be dynamically changing
- Assume that a collection of nodes large enough
- Do not consider power constraints
6Validation Platform
- Pioneer 2DX
- Laser (for obstacle avoidance)
- Compass
- Wireless ethernet
- PC-104 stack (Pentium 1Ghz)
- Motes
- Atmel ATmega 128L
- Program Flash Memory 128K bytes
- Measurement (Serial) Flash 512K bytes
- Configuration EEPROM 4 K bytes
- Serial Communications UART
- 916 MHz Multi-Channel Radio Transceiver
- Player/Stage engine
7Problems and solutions with Mobile Robots and
Sensor Networks
8Coverage and Exploration
- Deploy a group of robots maximizing sensor
coverage - Cover/Explore every point of environment with an
onboard sensor - Cant tell if coverage is complete
- Cant recover robots
- Require a LOT of expensive robots
- Use Sensor Network
9Approach
- Robot Loop
- If no beacon within radio range
- deploy beacon
- Else
- move in direction suggested by nearest beacon
- Beacon Loop
- Emit least recently visited direction
M. Batalin, G. S. Sukhatme, Coverage, Exploration
and Deployment by a Mobile Robot and
Communication Network, Telecommunications
Systems, April 2004 (accepted, to appear) M.
Batalin, G. S. Sukhatme, Efficient Exploration
Without LocalizationProceedings of the 2003 IEEE
International Conference on Robotics and
Automation (ICRA'03), Taipei, Taiwan, May 12 -
17, 2003.
10 Simulation Experiment
11Sensor Network Repair and Maintenance - Implicit
12Benefits
- A static sensor network is deployed
- can be used for numerous network applications
- guarantees that every point of environment would
be eventually covered by the mobile robots - etc.
- Robots can
- be used for exploration, patrolling and coverage
tasks - restore the static sensor network in case of
damage - retrieve robots
- system knows when full coverage is achieved
13Summary
- Theoretical analysis shows that trade offs in the
assumptions affect the outcome significantly - RW robot does not have or assume anything
RW(G) ?(VlogV ) - Proposed Approach the number of nodes is
enough ?(V ) and o(VlogV) - DFS nodes of three colors, perfect localization
and navigation DFS(G) T(V E) - Proposed algorithm outperforms other algorithms
if localization and perfect navigation cannot be
assumed
14Probabilistic Navigation
15Introduction
- How to navigate from point A to point B
- A fundamental problem in robotics
- No a priori information about the environment
- Use Sensor Network
16Transition Probabilities
- While robot deploys and traverses sensor network
from node to node it can determine transition
probabilities - Robot stores determined probabilities on
appropriate nodes
17Navigation Field computation
- Suppose transition probabilities are determined
for all nodes - If a goal node is specified, the information is
propagated through the network and the
navigation field is computed using - Where - V is the utility value, C(s, a) is the
cost associated with an action, P(ss, a) is the
transition probability of arriving to node s if
an action a was taken at node s, p(s) is the
policy, or direction that the node s is going to
suggest for the robot in the vicinity.
18Distributed Computation
- Suppose the SN is flooded with goal node data
- Every node updates own utilities according to
utility equation - After the utilities are computed, every node
computes an optimal policy for itself according
to policy equation - Note that the action policy computation is done
only once, and does not need to be recomputed,
unless the goal changes - Note that if neighbors of all nodes are known
exactly the system converges after a single
iteration.
19Basic algorithm
- Assume that SN deployed and Navigation Field is
computed - Closest Node suggests direction of motion
- Robot moves straight in a suggested direction
- Determine if close to the next node based on
signal strength - If no, repeat 2
- If yes, start 1
M. Batalin, G. S. Sukhatme, Coverage, Exploration
and Deployment by a Mobile Robot and
Communication Network, Telecommunications
Systems, April 2004 (accepted, to appear)
20Simulation Experiment
Node-to-node navigation Move in suggested
direction, switch to closest node, repeat
21Probabilistic Navigation in Real Environment
My Cube
22Real-World Navigation Challenges
- Cubicle environment is narrow, hence precision
is required - Compass or IMU proved to be useless inside
- Implementation should be simple enough for simple
robot - Algorithm should be based purely on signal
strength - Different antennas, not truly omnidirectional
- Ambient noise in the environment not constant
with time - Hence raw signal strength thresholding or an
observation model would not work reliably
23Basic algorithm
- Assume that SN deployed and Navigation Field is
computed - Robot knows its initial heading and closest node
- Closest Node suggests direction of motion
- Robot moves in a suggested direction
- Vector Field Histogram (VFH) algorithm is used
for local navigation and obstacle avoidance - Drives the robot from A to B avoiding obstacles
on the path on the local scale ( - Author J. Borenstein, available in Player/Stage
- Determine if close to the next node (next slide)
- If no, repeat 2
- If yes, start 1
M. Batalin, G. S. Sukhatme,M. Hattig, "Mobile
Robot Navigation using a Sensor Network, To
appear IEEE International Conference on Robotics
and Automation, 2004
24When to switch between the nodes?
- Similar to general state estimation problem
- Difficult problem in general
- Proposed algorithm which estimates when the robot
is close to a node - Compute initial maximum average of the first i
samples - Compute a running average which is an average of
j consecutive samples where j - Threshold on ratio of averages
25Experiments The Environment
My Cube
3
1
2
5
4
6
8
7
9
26Experiments Run 1
1
9
27Summary
- Algorithm allows the robot to navigate precisely
and reliably using a deployed sensor network. - Approach differs from systems described in the
literature by assuming that the map,
localization, GPS, IMU and compass are not
available - The navigation occurs through node-wise motion
from node to node on the path from starting node
to the goal node - We conducted 50 experiments for 5 different
goals, totaling in over 1 km of traveled
distance - In each of the 50 cases the robot successfully
navigated to the goal node - Note that we considered an experiment to be
successful if the robot approached the goal node
to within 3m
28Task Allocation Problem
29Introduction
- Task Allocation (TA) is the problem of assigning
resources (robots for example) to tasks - Offline TA is the problem of assigning resources
to different tasks (processes) if the tasks
arrival distribution is known a priori - Task assignments are computed offline
- Resembles Traveling Salesperson Problem, it is
NP-Complete - Online TA is the problem of assigning resources
to tasks if the distribution of the tasks
arrival is NOT known a priori - The task assignment occurs in decision epochs.
- A decision epoch is a period of time during which
only arrived tasks are considered for assignment.
- It is shown in literature that in case of Online
TA the optimal solution assigns the tasks in a
greedy fashion - We consider an Online TA
- Note that many real life TA problems are Online
30Experimental Scenario
- We study a particular experimental scenario -
emergency handling - Alarms are detected by nodes in the static
network - The problem is to assign and navigate robots to
different alarms - The goal is to minimize the cumulative alarm
OnTime across all alarms, over the duration of
the entire experiment - Alarms OnTime is a difference between the time
the alarm was turned off by a robot and the time
the alarm was detected by one of the nodes of the
network
31Distributed In-Network Task Allocation (implicit)
- Suppose sensor network monitors the environment
- If an event is detected by node n it sends out a
packet (n_id, weight, hop_count) - Compute Navigation Field
- This computation results in a direction which
maximizes the net utility of the robot - If there are several events detected at the same
time, a node computes direction towards the goal
node with largest
32Multi Field Distributed In-Network Task
Allocation (explicit)
- If an event is detected by node n it sends out a
packet (n_id, time_of_event, weight, hop_count) - Every node considers task assignment in decision
epochs - At the end of current decision epoch, network
synchronizes current positions of available
robots - Since every node receives the same data the
node states are in synch an optimal greedy
assignment is possible. - Hence, every node maintains a suggested direction
per task
33Experimental results
- Player/Stage simulations
- Sensor Network of 25 motes
- Groups of 1-4 robots, 10 trials/group
- 10 Alarms are drawn from the Poisson distribution
with ?1/60 - Empty environment, A 576 m2
34Sensor Network Repair and Maintenance - Explicit
35Benefits of DINTA DINTA-MF compared to other
techniques
- Sensing, computation and communication are
distributed - Provides sensor that is everywhere at the same
time - Can estimate utilities directly
- Does not rely on global information (no map, GPS,
localization, etc.) - Can be combined with other techniques to increase
range of applications
M. Batalin, G. S. Sukhatme, Using a Sensor
Network for Distributed Multi-Robot Task
Allocation, To appearIEEE International
Conference on Robotics and Automation, 2004 M.
Batalin, G. S. Sukhatme, "Sensor Network-based
Multi-Robot Task Allocation," In IEEE/RSJ
International Conference on Intelligent Robots
and Systems, pp. 1939-1944, 2003
36Current Work Task Allocation for NIMS
3747 m
Solar Cell
91
NIMS Node
84
Power Distribution Cable
Imager With Pan/Tilt Actuator
50 m
Vertical Node (Ta, RH, PAR)
Battery Pack
38(No Transcript)
39Problem Definition
- Assume a NIMS system (s1, s2, s3, ) and a Sensor
Network deployed in the same area - Suppose nodes of the sensor network detect
phenomena (p1, p2, p3, ) that require further
study by a NIMS system - Compute an optimal assignment of (s1, s2, s3,
) (p1, p2, p3, ) - The problem is an instance of an Online Task
Allocation problem
40Current Status
- Implemented a version of the algorithm in
simulation - Porting to the lab NIMS system a model of an
actual node - In march plan experiments in James Reserve
41(No Transcript)
42Summary
- Coverage and Exploration through Sensor Network
Deployment - Sensor Network Repair and Maintenance Implicit
and Explicit - Probabilistic Navigation
- Task Allocation Problem (scheduling)
- DINTA
- DINTA-MF
- Current Work Task Allocation for NIMS
Symbiotic systems are beneficial and important to
study
43Contacts
- Maxim A. Batalin
- Robotic Embedded Systems Laboratory
- Center for Robotics and Embedded Systems
- Computer Science Department
- University of Southern California
- Los Angeles, CA 90089, USA
- maxim_at_robotics.usc.edu
- http//www-robotics.usc.edu/maxim