COMMUNICATION IN MULTIROBOT TEAMS

- BARATH CHRISTOPHER PETIT

PAPER

COMMUNICATION IN REACTIVE MULTIAGENT SYSTEMS -Tucker Balch and Ronald Arkin

ISSUES

● Multiple robot teams are faster and reliable than a single robot system.

● Can communication between the robots in a multirobot team enhance the performance?

● What level of communication yields the best performance?(relative to the performance metrics being used to evaluate the performance.)

PERFORMANCE EVALUATION

● Three benchmark tasks were devised to evaluate system performance, namely:

● FORAGE● CONSUME● GRAZING

PERFORMANCE EVALUATION

● Six parameters for classification● TASK● COMMUNICATION TYPE● NUMBER OF ROBOTS● NUMBER OF ATTRACTORS● MASS OF ATTRACTORS● PERCENTAGE OF OBSTACLE COVERAGE

THE FORAGE TASK

● Robot wanders in environment looking for items of interest (attractors).

● Once attractor is sighted, robot moves towards it, acquires it and finally returns it to specified home base.

● Mass of attractor dictates completion time.● Several robots can cooperate in carrying an attractor to

home base but speed of slowest robot will be the bottleneck.

THE CONSUME TASK

● Similar to FORAGE but after acquiring the attractor the robot operates on the attractor instead of carrying it to a home base.

● Time to completion is proportional to mass of attractor.● Several robots can cooperate while operating or

'consuming' the attractor.● Rate of consumption is linear with number of operating

robots, (there is no ceiling).

THE GRAZE TASK

● Unlike FORAGE, CONSUME there are no discrete attractors.

● Aim is to completely visit the environment (or some percentage of it).

● Time of completion dictated by size of the environment.● Multiple robots reduce time if they avoid previously

grazed areas and if they can sight ungrazed areas quickly.

TASK PARAMETERS

● NUMBER OF ATTRACTORS (for FORAGE and CONSUME).

● MASS OF ATTRACTORS (for FORAGE and CONSUME).

● GRAZE COVERAGE (for GRAZE).

COMPLEX TASKS

Complex tasks can be viewed as being a combination of simpler tasks like

foraging, consume or grazing.

Wander Acquire

Deliver

Encounter

DepositAttach

The FORAGE FSA

THE WANDER STATE

● Noise: high gain to maximize coverage.● Avoid-static-obstacle (for objects): high to avoid

collisions.● Avoid-static-obstacle (for robots) : high to avoid other

robots to ensure maximum and efficient area coverage.● Detect-attractor: a perceptual schema that is triggered

when an attractor is sighted, enabling the transition to the ACQUIRE state.

THE ACQUIRE STATE

● Noise: low gain, to overcome local minimas● Avoid-static-obstacle (for objects): high to avoid

collisions with obstacles.● Avoid-static-obstacle (for robots): very low so that

robots can converge on same attractor to cooperate.● Move-to-goal: high to move to the detected attractor.● Detect-attachment: a perceptual schema that is triggered

when robot is close enough to attach to the attractor.

THE DELIVER STATE

● Noise: low to overcome local minima.● Avoid-static-obstacle (for objects): high to avoid

collisions.● Avoid-static-obstacle (for robots): low to enable robots

to cooperate.● Move-to-goal: high (target is home base).● Detect-deposit: a perceptual schema that is triggered

when home base is reached.

Wander Acquire

Consume

Encounter

CompleteAttach

The CONSUME FSA

CONSUME

● WANDER and ACQUIRE states are similar to states in FORAGE but instead of DELIVER , the CONSUME state is used.

● In CONSUME, only one motor schema is active which reduces the mass of the attractor till it becomes zero.

● Once attractor is consumed, the robot transitions to the WANDER state.

Wander Acquire

Graze

Encounter

Encounter grazed area Move to ungrazed part

The GRAZE FSA

The GRAZE state

● Noise: low gain to overcome local minima.● Avoid-static-obstacle: (for objects) high to avoid

collision.● Avoid-static-obstacle: (for robots) very low to enable

robots to graze closeby.● Probe: moderate to enable robot to move along its

current heading.● Graze: the graze mechanism, which 'tags' the area being

grazed.● Detect-grazed-area: this triggers transition to WANDER.

INTER–AGENT COMMUNICATION

● NO COMMUNICATION● STATE COMMUNICATION● GOAL COMMUNICATION

NO COMMUNICATION

● Robots can sense other robots, obstacles and attractors.● However none of the information is transmitted to other

robots.● Note that robots can still cooperate (in tasks like foraging

and consume), as was shown by Arkin.

STATE COMMUNICATION

● Robots can detect internal states of other robots.● Single bit communication: a one indicates that the robot

is in WANDER state, and a zero indicates a state other than WANDER.

● State communication need not be deliberate.

GOAL COMMUNICATION

● Transmission and reception of goal information.● Needs to be deliberate, unlike previous two modes of

communication.● Robot can get to goal directly rather than follow the

sender robot (unlike state communication).

PERFORMANCE METRICS

● COST: minimize system cost. Aims to reduce number of robots.

● TIME: minimizes completion time. Tends to increase number of robots.

● ENERGY: minimizes the energy expended in completing the task.

● RELIABILITY/SURVIVABILITY: Priority is on completion of task.

BASELINE PERFORMANCE

● 3-dimensional plot● <number of robots, number of attractors, value of

metric>● Maximum time for single robot.● Minimum time for maximum robots.● In some cases, improvement in time with addition of

robots is not significant.● For given number of robots, it takes longer to complete

the task with more number of attractors.

AMORTIZED COST METRIC

● If systems needs to be both fast and inexpensive, then thers is a tradeoff.

● In this case the metric is: N*300 + T.● N: number of robots.● T: Time taken.● The cost of each robot per run is taken as (for eg.) 300.● For FORAGE, a system with 2 robots is best for 3-4

attractors.

SPEEDUP

● Measures efficiency of n robot team relative to a single robot system in performing a task.

● If attractors are fixed, then it is defined as : (time taken by 1 robot)/ (n * time taken by n robots).

● Speedup is higher for larger number of attractors.● Speedup is sublinear for CONSUME, but can be

superlinear for lower mass of attractors.● Speedup for GRAZE is mostly superlinear.

RESULTS WITH COMMUNICATION

● FORAGE:State communication improved performance by 16%. Goal communication is better than state communication by 3%.

● CONSUME: State communication gives 10% improvement. Goal communication gives 6% improvement (over no communication).

● For low mass attractors, Goal communication almost indistinguishable from state communication.

● GRAZE: Communication has hardly any effect, due to implicit communication.

RESULTS

● Initial testing on mobile robots support simulation results.

● Comunication improves performance in tasks with little implicit communication.

● Communication is not effective in tasks which include implicit communication.

● More complex strategies offer little or no benefit over low-level communication.