Why Robots may need to be self-‐aware, before we can really trust them - Alan Winfield.
0 likes733 views
The document discusses how robots may need to be self-aware to be trusted, especially in unpredictable environments. It argues that safety cannot be achieved without self-awareness when a robot's environment is unknown. An internal model allows a robot to simulate possible future actions and outcomes without committing to them. This can provide a minimal level of functional self-awareness for safety. A generic internal modeling architecture is proposed where an internal model evaluates consequences of actions to moderate action selection for safety. Examples of robots using internal models for functions like planning, learning control, and distributed coordination are also provided.
Why Robots may need to be self-‐aware, before we can really trust them - Alan Winfield.
- 1. Why Robots may need to be self-‐aware, before we can really trust them Alan FT Winfield Bristol Robo=cs Laboratory Awareness Summer School, Lucca 26 June 2013
- 2. Outline • The safety problem • The central proposi=on of this talk • Introducing Internal Models in robo=cs • A generic Internal Modelling architecture, for safety – worked example: a scenario with safety hazards • Towards an ethical robot – worked example: a hazardous scenario with a human and a robot • The major challenges • How self-‐aware would the robot be? • A hint of neuroscien=fic plausibility
- 3. The safety problem • For any engineered system to be trusted, it must be safe – We already have many examples of complex engineered systems that are trusted; passenger airliners, for instance – These systems are trusted because they are designed, built, verified and operated to very stringent design and safety standards – The same will need to apply to autonomous systems
- 4. The safety problem • The problem of safe autonomous systems in unstructured or unpredictable environments, i.e. – robots designed to share human workspaces and physically interact with humans must be safe, – yet guaranteeing safe behaviour is extremely difficult because the robot’s human-‐centred working environment is, by defini5on, unpredictable – it becomes even more difficult if the robot is also capable of learning or adapta5on
- 5. The proposi=on In unknown or unpredictable environments, safety cannot be achieved without self-‐awareness
- 6. What is an internal model? • It is an internal mechanism for represen=ng both the system itself and its environment – example: a robot with a simula5on of itself and its currently perceived environment, inside itself • The mechanism might be centralized, distributed, or emergent “..an internal model allows a system to look ahead to the future consequences of current ac=ons, without actually commiYng itself to those ac=ons” John Holland (1992), Complex Adap=ve Systems, Daedalus.
- 7. Using internal models • Internal models can provide a minimal level of func5onal self-‐awareness – sufficient to allow complex systems to ask what-‐if ques=ons about the consequences of their next possible ac=ons, for safety • Following Dennea an internal model can generate and test what-‐if hypotheses: – what if I carry out action x..?! – of several possible next actions xi, which should I choose?!
- 8. Dennea’s Tower of Generate and Test Darwinian Creatures Skinnerian Creatures Popperian Creatures Dennea, D. (1995). Darwin’s Dangerous Idea, London, Penguin. Natural Selec=on Individual (Reinforcement) Learning Internal Modelling
- 9. Examples 1 • A robot using self-‐ simula=on to plan a safe route with incomplete knowledge Vaughan, R. T. and Zuluaga, M. (2006). Use your illusion: Sensorimotor self-‐ simula=on allows complex agents to plan with incomplete self-‐knowledge, in Proceedings of the Interna=onal Conference on Simula=on of Adap=ve Behaviour (SAB), pp. 298–309.
- 10. Examples 2 • A robot with an internal model that can learn how to control itself Bongard, J., Zykov, V., Lipson, H. (2006) Resilient machines through con=nuous self-‐ modeling. Science, 314: 1118-‐1121.
- 11. Examples 3 • ECCE-‐Robot – A robot with a complex body uses an internal model as a ‘func=onal imagina=on’ Marques, H. and Holland, O. (2009). Architectures for func=onal imagina=on, Neurocompu=ng 72, 4-‐6, pp. 743–759. Diamond, A., Knight, R., Devereux, D. and Holland, O. (2012). Anthropomime=c robots: Concept, construc=on and modelling, Interna=onal Journal of Advanced Robo=c Systems 9, pp. 1–14.
- 12. Examples 4 • A distributed system in which each robot has an internal model of itself and the whole system – Robot controllers and the internal simulator are co-‐ evolved O’Dowd P, Winfield A and Studley M (2011), The Distributed Co-‐Evolu=on of an Embodied Simulator and Controller for Swarm Robot Behaviours, in Proc IEEE/RSJ Interna=onal Conference on Intelligent Robots and Systems (IROS 2011), San Francisco, September 2011.
- 13. A Generic IM Architecture for Safety Internal Model Evaluates the consequences of each possible next ac=on Sense data Actuator demands The loop of generate and test The IM is ini=alized to match the current real situa=on Robot Controller The IM moderates ac=on-‐selec=on in the controller Copyright © Alan Winfield 2013
- 14. A Generic IM Architecture for Safety Sense data Actuator demands The loop of generate and test Robot Controller Robot Controller Robot Model World Model Consequence Evaluator Object Tracker -‐ Localiser Copyright © Alan Winfield 2013
- 15. N-‐tuple of all possible ac=ons (a1, a2, a3, a4) A Generic IM Architecture for Safety Sense data Actuator demands The loop of generate and test Robot Controller Robot Controller Robot Model World Model Consequence Evaluator Object Tracker -‐ Localiser Copyright © Alan Winfield 2013
- 16. N-‐tuple of all possible ac=ons (a1, a2, a3, a4) A Generic IM Architecture for Safety Sense data Actuator demands The loop of generate and test Robot Controller Robot Controller Robot Model World Model Consequence Evaluator Object Tracker -‐ Localiser S-‐tuple of safe ac=ons (a3, a4) Copyright © Alan Winfield 2013
- 17. A Generic IM Architecture for Safety Sense data Actuator demands The loop of generate and test Robot Controller Robot Controller Robot Model World Model Consequence Evaluator Object Tracker -‐ Localiser S-‐tuple of safe ac=ons (a3, a4) N-‐tuple of all possible ac=ons (a1, a2, a3, a4) Copyright © Alan Winfield 2013
- 18. A scenario with safety hazards Consider a robot that has four possible next ac=ons: 1. turn leq 2. move ahead 3. turn right 4. stand s=ll Hole Robot Wall Copyright © Alan Winfield 2013
- 19. A scenario with safety hazards Consider a robot that has four possible next ac=ons: 1. turn leq 2. move ahead 3. turn right 4. stand s=ll Hole Wall Robot ac(on Posi(on change Robot outcome Consequence Ahead leq 5 cm Collision Robot collides with wall Ahead 10 cm Collision Robot falls into hole Ahead right 20 cm No-‐collision Robot safe Stand s=ll 0 cm No-‐collision Robot safe Copyright © Alan Winfield 2013
- 20. Towards an ethical robot Which robot ac=on would lead to the least harm to the human? Copyright © Alan Winfield 2013
- 21. Towards an ethical robot Which robot ac=on would lead to the least harm to the human? Robot ac(on Robot outcome Human outcome Consequence Ahead leq 0 10 Robot safe; human falls into hole Ahead 10 10 Both robot and human fall into hole Ahead right 4 4 Robot collides with human Stand s=ll 0 10 Robot safe; human falls into hole Outcome scale 0:10, equivalent to Completely safe: Very dangerous Copyright © Alan Winfield 2013
- 22. Combining safety and ethical rules IF for all robot actions, the human is equally safe! THEN (* default safe actions *)! !output s-tuple of safe actions! ELSE (* ethical actions *)! !output s-tuple of actions for least unsafe human outcomes! Consider Asimov’s 1st and 3rd laws of robo=cs: (1) A robot may not injure a human being or, through inac=on, allow a human being to come to harm, (3) A robot must protect its own existence as long as such protec=on does not conflict with the First (or Second) Laws Isaac Asimov, I, ROBOT, 1950 Copyright © Alan Winfield 2013
- 23. Extending into Adap=vity Sense data Actuator demands The loop of generate and test Robot Controller Robot Controller Robot Model World Model Consequence Evaluator Object Tracker -‐ Localiser Learned/adap=ve behaviours Copyright © Alan Winfield 2013
- 24. Extending into Adap=vity Sense data Actuator demands The loop of generate and test Robot Controller Robot Controller Robot Model World Model Consequence Evaluator Object Tracker -‐ Localiser Learned/adap=ve behaviours Copyright © Alan Winfield 2013
- 25. Challenges and open ques=ons • Fidelity: to model both the system and its environment with sufficient fidelity; • To connect the IM with the system’s real sensors and actuators (or equivalent); • Timing and data flows: to synchronize the internal model with both changing perceptual data, and efferent actuator data; • Valida5on, i.e. of the consequence rules.
- 26. Major challenges: performance • Example – imagine placing this Webots simula=on inside each NAO robot: Note the simulated robot’s eye view of it’s world
- 27. A science of simula=on: the CoSMoS approach The Complex Systems Modelling and Simula=on (CoSMoS) process, from Susan Stepney, et al, Engineering Simula=ons as Scien=fic Instruments — a paaern language, Springer, in prepara=on. The CoSMoS Process Version 0.1: A Process for the Modelling and Simula=on of Complex Systems, Paul S. Andrews, et al, Dept of Computer Science, University of York, Number YCS-‐2010-‐453
- 28. Major challenges: =ming • When and how oqen do we need to ini=ate the generate-‐and-‐test-‐loop (IM cycle)? – Maybe when the object tracker senses a nearby object star=ng to move..? • How far ahead should the IM simulate – Let us call this =me ts. if ts is too short the IM will not encounter the hazard; too long will slow down the robot. – Ideally ts and its upper limit should be adap=ve.
- 29. How self-‐aware would this robot be? • The robot would not pass the mirror test – Haikkonen (2007), Reflec=ons of consciousness • However, I argue this robot would be minimally but sufficiently self-‐aware to merit the label – But this would have to be demonstrated by the robot behaving in interes5ng ways, that were not pre-‐programmed, in response to novel situa5ons – Valida=ng any claims to self-‐awareness would be very challenging
- 30. Some neuroscien=fic plausibility? • Libet’s famous experimental result showed that ini=a=on of ac=on occurs before the conscious decision to make take that ac=on – Libet, B (1985), Unconscious cerebral Ini=a=ve and the role of conscious will in voluntary ac=on, Behavioral and Brain Science, 8, 529-‐539. • Although controversial there appears to be a growing body of opinion toward consciousness as a mechanism for vetoing ac=ons – Libet coined the term: free won’t
- 31. In conclusion • I strongly suspect that self-‐awareness via internal models might prove to be the only way to guarantee safety in robots, and by extension autonomous systems, in unknown and unpredictable environments – and just maybe provide ethical behaviours too Thank you! Reference for the work of this talk: Winfield AFT, Robots with Internal Models: A Route to Self-‐Aware and Hence Safer Robots, accepted for The Computer AJer Me, eds. Jeremy Pia and Julia Schaumeier, Imperial College Press, 2013.