Biological Inspirations for Distributed Robotics Dr. Daisy Tang
Outline Biological inspirations Understand two types of biological parallels Understand key ideas for distributed robotics obtained from study of biological systems Understand concept of stigmergy Understand use of stigmergy for tasks in collective robotics
Biology vs. Multi-Robot Teams
Movies of Some Animal Collectives School of fish http://www.youtube.com/watch?v=_tgokngtkt4&feature=related Flock of birds http://www.youtube.com/watch?v=tl8dih-i9eq Etc.
Why Biological Systems? Key reasons: Animal behavior defines intelligence Animal behavior provides existence proof that intelligence is achievable Typical objects of study: Ants Bees Birds Fish Herding animals
A Broad Classification of Animal Societies (Tinbergen, 1953)
Societies that Differentiate Innate differentiation of blood relatives Strict division of work and social interaction Individuals: Exist for the good of society Are totally dependent on society Examples: Bees Ants, termites Stay Together
A Typical Bee Colony
Societies that Integrate Depend on the attraction of individual animals Exhibit loose division of labor Individuals: Integrate ways of behavior Thrive on support provided by society Are motivated by selfish interests Examples: Wolf, hunting dogs, etc. Bird colonies Come Together
Parallels to Cooperative Robotics
Which Approach To Choose? Differentiating approach: For tasks that require numerous repetitions of same activity over a fairly large area Examples: Waxing floor Removing barnacles off ships Collecting rock samples on Mars Integrating approach: For tasks that require several distinct subtasks Examples: Search and rescue Security, surveillance, or reconnaissance
Key Ideas from Biological Inspiration Communication Auditory, chemical, tactile, electrical Direct, indirect, explicit, implicit Roles Strict division vs. loose assignments Hierarchies Absolute linear ordering, partial ordering, relative ordering Purpose: reduction in fighting, efficiency Territoriality Reduces fighting, disperses group, simplifies interactions Imitation Complex mechanism for learning
Our Distributed Robotics Studies First: low-level, homogeneous, swarm robots Swarming, dispersion, homing, etc. Search/coverage Etc. Then: higher-level strategies, heterogeneous robots Multi-robot path planning, traffic management Task allocation Etc.
Key Concept in Swarm Distributed Robotics: Stigmergy Stigmergy: Term used by some biologists to describe influence on behavior due to persisting environmental effects of previous behavior Originally used by French biologist Pierre-Paul Grasse to describe behavior of nest-building termites and trails Equivalent concept: implicit communication by means of modifying the environment A mechanism for binding task state information to local features of a task site, and for communicating (implicitly) by modifying those features Stigmergy is a powerful tool for coordination in a loosely coupled system
Stigmergy in Nature Ant trails Ants find the shortest path to a food source in their vicinity using stigmergy to maintain traffic statistics Termite nest-building Termites build columns and arches using stigmergy to retain state about the building process Ant corpse-gathering Ants pick up dead ants and drop them in piles, preferring larger piles, until there is only one pile left
Ants Finding The Shortest Path Ants follow random paths, influenced by presence of pheromones Ants returning with food leave stronger trails Pheromones evaporate, causing frequent trails to dominate Shortcuts result in higher traffic (more trips per ant per unit time) and thus are selected with greater probability http://www.youtube.com/watch?v=kn0m49iqfrc
Termites Building An Arch Termites make mud balls with pheromones Termites deposit mud balls near existing pheromone concentrations As columns get taller pheromones on the bottom evaporate Pheromones on neighboring columns cause the top to be built together to form an arch http://www.youtube.com/watch?v=0m7odgafpqu&feature=pl aylist&p=598428ddc4e49d85&index=0&playnext=1
Ant Corpse-Gathering Scattered corpses are picked up and dropped Small piles form Gradually the piles are aggregated into a single large pile
How Does Stigmergy Produce Complex Patterns? The state of the environment, and the current distribution of agentswithin it, determine how the environment and the distribution of agents will change in the future Any structure emerging is a result of selforganization Self-organization: A set of dynamic mechanisms whereby structures appear at global level of a system resulting from interactions among lower-level components Rules specifying interaction are executed purely based on local information, without reference to global pattern
Minimal Qualities of Agent and Environment to Support Stigmergy Agent has 2 key abilities: It can move through environment It can act on environment To enable stigmergy: Environment must be changed locally by agents Changes must persist long enoughto affect the choice, parameters, or consequences of agents behavior
Two Ways to Structure Behavioral Sequences In solitary species Execution of first movement in sequences sets internal state With external cue, internal state initiates second movement, etc. In solitary and social insects No internal state required (in many, but not all, cases) External cue alone is sufficient to invoke subsequent actions Sets the stage for stigmergy
Compare Stigmergy to Direct Communication Direct communication requires: Sending robot to encode and transmit message about what is to be done, and where it is to be done Implies knowledge of location Message is local in time and space, therefore only robots close enough and not otherwise engaged will be free to receive the message Robots must decode received messages, and remember them long enough to get to the place and carry out the action, or even longer if they are currently carrying out a more important task Stigmergic communication: Requires no encoding or decoding Requires no knowledge of place Requires no memory Is not transient
Example Use of Stigmergy in Collective Robotics Paper references: Stigmergy, Self-Organization, and Sorting in Collective Robotics, by Holland and Melhuish, Artificial Life 5: 173-202, 1999. From Local Actions to Global Tasks: Stigmergy and Collective Robotics, by Beckers, Holland and Deneubourg in Brooks and Maes (eds.), Artificial Live IV: 181-189, Cambridge, MA: MIT Press, 1994.
Collective Pile Formation Task The robots ~20cm square base with two wheels and a gripper Battery powered Infrared (IR) sensors for obstacle detection Gripper force sensor Environment Square arena, about 2.5x2.5m 81 circular pucks (4cm) arranged on a 25cm grid Beckers approach
The Pile Formation Experiment The initial task given the robots was to push all the pucks into a single pile At the start of an experiment, robots are in the center, oriented randomly After each 10 minute interval, the robots are stopped and sizes and positions of clusters noted Experiment ends when all pucks are in single cluster
Robot Behaviors Very simple set of 3 behaviors: If IR sensor active: turn awayfrom obstacle through a random angle If force sensor active: Force sensor triggered when 3 or more pucks are pushed When sensor activates, pucks are dropped Reverse both motors for one second Then turn awayto a random angle Default: move forward until sensor activated
Back to Experiment (Becker) How it works? Robots move around randomly If they bump into a puck, they will push it along When they bump into their third puck, they drop Initially, all piles are of size 1 Robots will pick them up and will not deposit until they have collected 3 pucks A pile of 3 or more tends to get bigger Robots that hit a pile of 3 or more head-on will add their pucks to pile
How do Piles Aggregate? Initially, a few small clusters form quickly Then, gradually those clusters are aggregated This occurs when pucks are stripped from the edge of a pile and then deposited elsewhere Large piles have a larger ratio of areas in the middle to those on the edge. Therefore probability of hitting tangent to pile decreases with increasing pile size Thus larger piles have a larger probability of increasing as a result of this process
Where is the Stigmergy? By adding pucks to a pile, a robot makes the pile larger, and votes (implicitly) for that pile to be largest This stigmergically encodes a message this is the largest pile, add more pucks to it The strongest such message (i.e. the largest pile) wins and eventually accretes all the pucks Because all state information is encoded in observed pile size, new robots can be added with no communication overhead
Experimental Results (Beckers) The experiment was performed with varying numbers of robots Adding robots sped convergence, up to 3 robots Why? More than three robots got in each others way (i.e., interference) Whenever they turn to avoid each other, they run the risk of scattering a nearby pile as they turn away Because the frequency of interactions increases with more robots, 3 was experimentally determined to be optimal Interference is a function of robot density
Experimental Results (Cont.) 1. Over time, size of the biggest cluster grows 2. More robots faster cluster growth up to a point of robot interference 1. Over time, # of clusters shrinks 2. More robots faster reduction, up to a point of robot interference
Experimental Results (Cont.) For these experiments, 3 robots was optimal Number of interactions increased significantly with number of robots Robot efficiency for these experiments was optimized at 3 robots
Summary of Stigmergy Stigmergy piggybacks communication on top of robot s existing sensing and actuation Allows system to scale to additional robots with additional communication overhead Although high densities can lead to gridlock, etc. Stigmergy stores state in the environment so that it is easily retrieved by specialized sensors In nature, pheromones In robotics, variety of sensors Stigmergy can be regarded as the general exploitation of the environment as external memory resource
Second Case Study Title Multi-robot System Based on Model of Wolf Hunting Behavior to Emulate Wolf and Elk Interactions Authors: John D. Madden, Ronald C. Arkin and Daniel R. MacNulty IEEE International Conference on Robotics and Biomimetics, 2010
Goal of the Project Models of behavior from biology are used to develop heterogeneous unmanned network teams (HUNT) The ability to reduce communication and planning requirements for robot groups, while still achieving missions Mission: pursuit-evasion tasks
Wolf Behavior from Nature No obvious pattern of coordinated hunting behavior Rules of thumb: Attack while minimizing risk of injury with no overall had behavioral constraints on actions
Transitions Model
Other Transitions are Possible Statistical observational data of state transitions
Coordination or Lack Thereof Wolves show no signs of planned strategies and no noticeable communication while hunting They do not make transitions together Coordination is a byproduct where each individual is maximizing its own utility Seeing elk being chased signals a sign of weakness of the prey, so they join the pursuit
Implementation of Wolf Behaviors with MissionLab @GaTech
List of Releasers and Transitions Weighted roulette wheel was used to decide which transition to take
List of Behaviors
List of Behaviors, Cont d
Elk Behaviors
Simulation Results
Transition Results Similar to Observed Data
Conclusion High fidelity bio models can provide utility for a range of multi-robot applications Byproduct mutualism can provide robust results for bio groups The ability to reduce communication and planning for robot groups
Summary of Biological Inspirations Study social biological systems either to: Obtain inspiration for how to build multi-robot systems Validate theoretical models for how biological systems work Two types of biological parallels: differentiating and integrative Many possible sources of inspiration from biology Stigmergy is important concept for swarm cooperation through the world