Adaptive Robotics COM2110 Autumn Semester 2008 Lecturer: Amanda Sharkey 1 So far … Lect 1: what is a robot? Brief history of robotics Early robots, Shakey and GOFAI, Behaviour-based robotics Mechanisms and robot control (and biological inspiration) Lect 2: Grey Walter, Brooks and Subsumption Architecture. Lect 3: Adaptation and learning Lect 4: Artificial Neural Nets and Learning Lect 5: Evolutionary Robotics 2 “Robots in the news” Honda – new wearable assisted walking gadget Designed to support bodyweight, reduce stress on the knees and help people get up steps and stay in crouching positions. To be used by workers in auto factories To be tested next month with assembly-line workers Based on technology developed for their Asimo robot 3 Phoenix: NASA Martian probe Has come to the end of its mission Not enough light to recharge batteries, and winter Has been on Mars for 5 months – sent back 25000 images 4 Due by Monday 17th Nov at 11 am Write an essay (1500-2500 words) on one of the following topics. You should use the lectures as a starting point, but also research the topic yourself. Plan your answer. Include a reference section, with the references cited in full. 1. Identify the main characteristics of Behaviour-based robotics, and contrast the approach to that of “Good old-fashioned AI”. 2. To what extent did Grey Walter’s robots, Elsie and Elmer, differ from robots that preceded, or followed them. 3. Explain how the concepts of “emergence” and “embodiment” are related to recent developments in robotics and artificial intelligence. 5 Collective Robotics Swarm Robotics 6 Collective robotics Why invest in collections of robots, why not build a reliable individual robot? - Task difficult (or impossible) for one robot Can be performed better by many Redundancy – task more likely to be completed Simplicity – many cheaper robots instead of one expensive one. 7 What kinds of collections? Possibilities range from - remote controlled robots - centrally controlled robots - completely autonomous robots 8 Cao, Fukunaga and Kahn (1997) Cooperative mobile robotics: antecedents and directions. Autonomous Robots, 4,1, 7-27. 9 Advantages of robot collectives shown in Environmental exploration Materials transport Coordinated sensing – collective cooperates to provide maximal sensor coverage of moving target. Robot soccer Search and Rescue 10 Swarm Robotics Taking a swarm intelligence approach to robotics 11 Swarm intelligence Swarm intelligence is “any attempt to design algorithms or distributed problem-solving devices inspired by the collective behaviour of social insect colonies and other animal societies” Bonabeau, Dorigo and Theraulaz (1999) 12 13 Natural swarms Decentralised – no-one in control Individuals are simple and autonomous Local communication and control Cooperative behaviours emerge through self- organisation e.g. repairing damage to nest, foraging for food, caring for brood 14 15 Self-organisation Organisation increases in complexity, without external guidance Self-organising systems often display emergent properties “self-organisation is a set of dynamical mechanisms whereby structures appear at the global level of a system from interactions among its lower-level components. The rules specifying the interactions among the system’s constituent units are executed on the basis of purely local information, without reference to the global pattern, which is an emergent property of the system rather than a property imposed upon the system by an external ordering influence” (Bonabeau, Dorigo and Theraulaz, 1999) 16 Emergence An emergent property, e.g. pattern formation, from more basic constituents An emergent behaviour can appear as a result of the interaction of components of the system E.g. flocking, or organisation of ant colony 17 Real life example of self-organised behaviour in humans Emergence of paths across grassy area Most popular paths are reinforced Counter –example e.g. a team of carpenters building a house….not self-organised. 18 Swarm robotics Inspired by self-organisation of social insects Using local methods of control and communication Local control: autonomous operation Local communication: avoids bottlenecks Scalable – new robots can be added, or fail without need for recalibration Simplicity – cheap, expendable robots Self-organisation Decentralisation 19 Disadvantages of centralised control and communication. Central control: failure of controller implies failure of whole system Robot to robot communication becomes very complex as number of robots increases. Communication bottlenecks Adding new robots means changing the communication and control system 20 Applications of swarm approach Some tasks are particularly suited to group of expendable simple robots e.g. - cleaning up toxic waste - exploring an unknown planet - pushing large objects - surveillance and other military applications 21 What issues are investigated? Weak AI questions: E.g. how can complex behaviour, such as cooperation, emerge as a result of interactions between simple agents and their environment? – Biological modelling – better understanding of social insects for example. - Biological inspiration – emulating behaviour and capabilities of biological systems 22 Cooperation and communication Examples of communication in cooperative systems: Increasing sophistication…. Bacteria Ants Wolves Non-human primates Humans 23 Bacteria Live in colonies Explicit chemical signals mediate their ability to cooperate. E.g. Mycobacteria assemble into multicellular structures known as fruiting bodies. Bacteria emit and react to chemical signals 24 Ants Also termites, bees and wasps Display cooperative behaviour e.g. pheromone trails to food source Chance variations that result in shorter trail are reinforced at faster rate. Can find optimal shortest path Stigmergic communication. 25 Wolves Territory marking through repeated urination on objects on periphery of territory Also more sophisticated communication directed at particular individuals Specific postures and vocalisations 26 Non-human primates Sophisticated cooperative behaviour Higher primates can represent the internal goals, plans, dispositions and intentions of others, and to construct collaborative plans jointly through acting socially. 27 Humans Many forms of communications – including written and spoken language Many forms of cooperation, from basic altruism to cooperative relationships where we exchange resources for mutual benefit 28 Focus of interest here: Emergent cooperation e.g. social insects: ants, bees, wasps, termites Stigmergic communication: one of the mechanisms that underlies cooperation. 29 Swarm robotics Biologically inspired by social insects - emergent complex behaviour from simple agents Swarm Intelligence Principles: Autonomous control Simple agents (debateable – swarms of helicopters?) Expendable, fast and flexible responses Local communication Scalable Decentralised Use and exploration of stigmergy 30 Mystery: cooperative behaviour when insects seem to work alone 31 32 33 individual insect responds to changes in environment created by itself or others Grassé (1959) – stigmergy - Indirect social interaction via the environment - 34 E.g. Termite nest building Building arches Termites make mudballs, which they deposit at random. Chemical trace added to each ball Termites prefer to drop mudballs where trace is strongest. Columns begin to form Deposit more on side nearest to next column – eventually leads to formation of arch. 35 Example paper: Holland and Melhuish (1999) Holland, O., and Melhuish, C., (1999) Stigmergy, self-organisation and sorting in collective robotics. Artificial Life, 5, 173-202. 36 37 Example of ant brood sorting “The eggs are arranged in a pile next to a pile of larvae and a further pile of cocoons, or else the three categories are placed in entirely different parts of the nest…if you tip the contents of a nest out onto a surface, very rapidly the workers will gather the brood into a place of shelter and then sort it into a different pile as before (Deneubourg, et al, 1991) 38 Franks and Sendova-Franks (1992) Brood sorting of Leptothorax unifasciatus - brood items sorted into concentric rings of progressively more widely spaced brood items at different stages of development. 39 Use of simulations Deneubourg et al (1991) “The dynamics of collective sorting: Robot-like ants and ant-like robots” Showed agents could use stigmergy to cluster scattered objects of a single type, and to sort objects of two different types For sorting – agents needed short-term memory to sense local density of different types of brood items and to know the type of brood item they were carrying. But – a simpler solution can be found with physical agents – greater exploitation of real world physics. 40 41 Holland and Melhuish experiments: Small U-bot robots, with infrared sensors, and gripper designed to sense, grip, retain, and release frisbees. When robot moves forward, frisbee remains in gripper When robot reverses, frisbee left behind, unless pin extended to keep it in place When 2 or more frisbees pushed into, this triggers microswitch in gripper – not triggered when pushing or bumping into 1 frisbee. 42 43 44 Exp 1: how many U-bots in arena without too many collisions Exp 2: Simple rule set Rule 1: if (gripper pressed and object ahead) then make random turn away from object -> ie turn away from boundary Rule 2: if (gripper pressed and no object ahead) then reverse small distance (dropping the frisbee) and make random turn left or right -> ie has encountered another frisbee. Rule 3: go forward 45 44 frisbees placed across the arena 10 robots released. Frisbees gradually collected in small clusters – after 8 hours 25 mins, a cluster of 40 frisbees formed. Frisbees taken from intermediate clusters if struck at an angle without triggering gripper 46 47 Experiment 5: sorting and pull-back algorithm Plain yellow frisbees and black and white ring frisbees Pin-dropping mechanism applied to plains Rule 1: if (gripper pressed and object ahead) then make random turn away from object Rule 2: if (gripper pressed and no object ahead) then If plain lower pin and reverse for pullback distance raise pin reverse small distance (dropping frisbee) make random turn left or right Rule 3: go forward 48 - now if robot is pushing a plain frisbee and hits another, or if not pushing frisbee and collides with another plain in a cluster, the plain will be dragged backwards and dropped away from contact point. Result (after 7h 35 m): central core of 17 ring frisbees with 11 plains and 4 rings round outside. I.e. annular sorting, based on simple mechanism - Example of seemingly complex behaviour (sorting) emerging from the application of simple rules. 49 50 51 What do these experiments show? Apparently co-operative behaviour, with no central control, and no direct communication. Segregation and crude annular sorting can be achieved by system of simple (reactive) mobile robots - robots can only sense the type of object they are carrying - they have no capacity for spatial orientation or memory 52 Some elements of these mechanisms found in social insects E.g. Leptothorax building behaviour: possible use of increased resistance to pushing a building block forward against other building blocks as a cue to drop it “the ants drop their granule if they meet sufficient resistance” (Franks et al, 1992) 53 Also mechanism like pull-back mechanism “workers individually carry granules into the nest. They walk head first towards the cluster of their nest-mates who are already installed in the nest, forming a fairly tight group. After coming close to the group of ants, the builder then turns through 180 degrees to face outwards from the nest. The worker then actively pushes the granule it is carrying into other granules already in the nest, or after a short time, if no other granules are encountered, it simply drops its load” (Franks et al, 1992) 54 Summary of Holland and Melhuish paper: A simpler solution obtained in the robotic experiments where the physics of the environment can be exploited, than in abstract computer simulations Simple behavioural rule set – no capacity for spatial orientation or memory, but robots able to achieve effective clustering and sorting Example of stigmergy – indirect communication via the environment. 55 Sorting and clustering accomplished here by robots with no memory, and no understanding of their task. Does this mean that ants also have no memory or understanding of their tasks? 56 Kube, C.R. and Bonabeau, E. (2000) Cooperative transport by ants and robots. Robotics and Autonomous Systems, 30, 85101. Current interest in robotics is result of Relative failure of classical AI program. Swarm-based robotics, and idea that group of robots can perform tasks without explicit representations of environment and other robots Mobile robots becoming cheaper and more efficient Artificial life and emphasis on emergent behaviour – increasing awareness of biological systems 57 “As the reader will perhaps be disappointed by the simplicity of the tasks performed by state-of-the-art robotic systems such as the one presented in this paper, let us remind her or him that it is in the perspective of miniaturisation that swarm-based robotics becomes meaningful … understanding the nature of coordination in groups of simple agents is a first step towards implementing useful multirobot systems” (Kube and Bonabeau, 2000) 58 Cooperative transport in robots, and cooperative prey retrieval in social insects E.g Moffett (1988) a group of 100 ants Pheidologeton diversus could transport a 10 cm earthworm weight 1.92g (more than 5000 x 0.3 mg minor worker) Single ants carry burdens at most 5 times their body weight 59 60 Questions about cooperative prey retrieval in social insects Is there an advantage to group transport over solitary transport? When and how does an ant know it cannot carry an item alone? How are nest mates recruited? How do several ants cooperate and coordinate their actions to transport an item? How do ants ensure the right number of ants help? How does a group of ants handle deadlocks? 61 Group vs solitary transport Moffet (1988) transport efficiency per ant (product of burden weight by transport velocity divided by no. of carriers) increases with group size up to a maximum for groups of 8-10, and then declines Switching from solitary transport resistance to transport 62 Recruitment of nestmates Holldobler et al (1978) African weave ant Aphaenogaster species Short range recruitment – releasing poison gland secretion in the air when prey discovered. Ants recruited from 2m distance Long range recruitment – chemical trail of poison gland recruitment laid from prey to nest. 63 Coordination in collective transport – not well understood. Movement of one ant engaged in group transport modifies the stimuli perceived by other group members (stigmergy) Number of ants – an increasing feature of how difficult (weight and resistance) it is to carry the prey Deadlock and stagnation recovery: ants show realigning and repositioning behaviours 64 Robot task: cooperative box pushing Previous versions – centralised planning and conflict resolution, with explicit communication between robots Kube and Zhang (1994) directed box pushing by robots Applied in simulation first, then robots 65 Non-directed box pushing Physical robots: 2 behaviours AVOID (left and right obstacle sensor mapped to left and right wheel motors) GOAL (left and right sensors mapped to right and left wheel motors causing robots to turn towards brightly lit box) Controllers allowed robots to locate box, converge and push it But stagnation could arise How do ants recover from stagnation? 66 Cooperative prey retrieval Sudd (1960) strategies to combat stagnation observed in ants cooperatively retrieving prey Realignment of body without releasing grasp If that fails, grasp released, and ants reposition Same strategies used for robot box pushing 67 Comparison of strategies 1. 2. 3. 4. No stagnation recovery Realignment only Reposition only Realignment and reposition Performance improved with strategies For small group strategy (2) best, for large group (3) is better, and (4) is best. 68 Directed box-pushing Now 3 phases Finding the box Moving towards the box If oriented with respect to goal, pushing box Box detection simplified by placing bright light on box Goal detection simplified by shining spotlight 69 70 Directed box-pushing Phase 1: robots execute FIND-BOX and MOVE-TO-BOX Phase 2: robots incorrectly positioned for pushing move counter clockwise round perimeter caused by cycling through FIND-BOX, MOVE-TO-BOX, and PUSH-TO-GOAL when contact is made. ?SEE-GOAL indicates robot on wrong side for pushing, and REPOSITION behaviour until empty position found. Phase 3: push to goal – robots continuously monitor ?SEE-GOAL. If robot cannot see goal it will reposition. 71 Varying the number of robots More robots = more interference 72 What do these experiments show? Coordinated group effort is possible without use of direct communication or robot differentiation - ants not always efficient – eg ants can take 10 minutes to begin moving object. Model makes testable predictions about stagnation recovery mechanism to be expected depending on ecological conditions and prey size E.g. adding mechanisms for stagnation recovery increases retrieval time, and probability of success Where little competition, should find more stagnation recovery mechanisms Where strong competition, should find less stagnation recovery 73 Research questions in Swarm Robotics Self-organised methods for task allocation to ensure that enough robots are allocated to a task Collective decision making Communication – local methods to detect when needs of group have changed Control and coordination of heterogeneous groups Incorporation of some learning and recognition – e.g. of landmarks 74 What have we looked at in this lecture? Idea of collective robotics Possible reasons for using several robots Swarm intelligence and swarm robotics Self-organisation and emergence Possible applications Cooperation and communication Forms of communication – Stigmergic indirect communication Example papers: Holland and Melhuish (1999) and sorting Kube and Bonabeau (2000) and box pushing 75