Socio-Cognitive Robot Architectures An Exploratory Overview 15-12-2010 Lorentz Centre HART Workshop work in progress Koen V. Hindriks Contact: k.v.hindriks@tudelft.nl 9-4-2015 Webpage : http://mmi.tudelft.nl/SocioCognitiveRobotics 1 Goal of this presentation • Collect your feedback about some preliminary ideas about designing / developing a socio-cognitive robot control architecture • I’d also like to collect some lessons learned based on your robot development experience; e.g. which pitfalls should be avoided. • Please jump in! I’d appreciate teamwork ;-) 2 Overview • Exploratory overview of cognitive robot control architectures • Basic Abstract Architecture Design • Summarizing: Current understanding of some key challenges 3 Towards Socio-Cognitive Robot Architectures • Challenge for cognitive architectures: real time autonomous processing needed to interact with dynamic world we live in. • Need for socio-cognitive architectures pushed by humanoid robots that interact with humans in a multi-modal fashion. • Towards an architecture for social interaction and teamwork • Klein, G., Woods, D. D., Bradshaw, J. M., Hoffman, R. R., & Feltovich, P. (2004). Ten challenges for making automation a "team player" in joint human-agent activity. IEEE Intelligent Systems 19(6): 91-95. • Here we look at various current state-of-the-art approaches, and take cognitive robot architectures as a starting point. 4 Cognitive Robot Control Architectures An Exploratory (and Necessarily Brief) Overview Delft University of Technology Challenge the future A Plethora of Architectures • • • • • • • • • • • • • • Subsumption architecture (Brooks 1985) BDL (Rochwerger et al. 1994) RAP (Firby 1994) TCA (Simons et al. 1997). SSS (Connell 1991) ATLANTIS (Gat 1991) 3T (Bonasso 1991) Saphira (Konolige 1996) CLARAty (Volpe et al 2001) CoSy schemas (Hawes et al 2007) Soar ACT-R (SS-RICS, …) ADAPT … 7 Stanford Cart Architecture Types Pipeline Architectures Based on a horizontal decomposition of functional components Vision Sensors Model Plan Robot Platform Execute Control Motors Environment • Classic architecture, also used for symbolic robot control architectures. • Potential to exploit parallelism, but hard and (typically?) not used in practice. 8 Architecture Types Behavior-Based Architectures Based on a vertical decomposition of behavior components Hannibal(MIT AI Lab) Behavior 4, e.g. Build Map Behavior 2, e.g. Avoid obstacle filter filter Behavior 3, e.g. Explore Behavior 1, e.g. Wander Sensors Robot Platform Motors Environment • Components are in competition, run in parallel and outputs are filtered by some technique. • Reactive architectures typically do not support cognitive functions and seem to have a “capability ceiling” (Gat 1998). 9 Alfred B12 Architecture Types 3T or Layered Architectures Based on a vertical decomposition of components Deliberator (High-level layer; planning, reasoning, …) Sequencer (Middle layer; conditional sequencing, sequencing constructs/language) Controller (Low-level layer; skills, feedback control loops) Sensors Robot Platform Motors Environment • Classic examples: SSS (Connell 1991), ATLANTIS (Gat 1991), 3T (Bonasso 1991) • High-level typically declarative techniques, low-level typically procedural techniques 10 Rationalizing 3T Architectures • Erann Gat (1998) rationalized three-layer architectures by arguing there is a correspondence between layers and the role of internal state. • Deliberator: state reflecting predictions about the future • Sequencer: state reflecting memories about the past • Controller: no state (stateless sensor-based algorithms) • Responsiveness, time scale also varies over components. 11 BIRON The Bielefeld Robot Companion (2004) 12 Care-O-bot 3 (Fraunhofer IPA, 2008) Instruction model Care-O-bot II/3 (FF) (MySQL) (JAM Agents) (Realtime Framework; RTF) 13 Armar Armar (Univ. of Karlsruhe) Low-level can also access GKB 14 Saphira Architecture “No overt planning” No third (high-level) layer LPS = Local Perceptual Space 15 CLARAty Architecture Two-layered architecture developed at JPL/NASA Observations: No standard no leverage of robotics’ community efforts Issues: “not invented here” “fear of unknown” “learning curve” … Observation: 3T: • dominant layer? • access to info? • obscures hierarchy within layers Two layers blend declarative and procedural techniques CLARA = Coupled Layered Architecture for Robotic Autonomy 17 B21r + Katana arm CoSy Architecture Schema Need for easy methods for linking modules using different forms of representation, without excessive run-time overhead integration mechanisms = architectural schema + binding information 19 Summarizing: Some key challenges Delft University of Technology Challenge the future Key Problem: Integration Challenge Observation: • Over time more and more components have been integrated into cognitive robot architectures. Q: • How many layers? • A Socio-Cognitive Architecture only adds to this challenge. Any ideas / approaches for effective design approaches for integrating e.g. new components for social interaction and coordination both with humans and other robots? 21 Key Problem: Access to Data/Information/KB Observation: • After classical 3T architectures, all cognitive robot architectures have a common database shared by all layers Q: • Which data needs to be shared? Mainly localization information? • It seems that all three-layered architectures require sharing of data by all layers. Do 2-layered architectures require this? 22 Well-defined Robot Architecture A well-defined architecture facilitates reuse and parallel development Q: • Do general software architectural principles apply? • What is a well-defined robot architecture? Any criteria? Example principles: • partition architecture into layers with well-defined interfaces • partition code into functional blocks with well-defined inputs and outputs •… 24 Basic Abstract Architecture Design Reducing the complexity? Delft University of Technology Challenge the future Abstract Architecture (1/2) Based on a vertical decomposition into functional layers Cognitive Layer P1 P2 Sensors … Behavioral Layer Robot Platform B1 B2 … Motors Environment • P1, P2, … = process 1, process 2, …; B1, B2, … = behavior 1, behavior 2, … • Cognitive functions supported in cognitive layer, e.g. reasoning, planning, memory, … 26 Abstract Architecture (2/2) Simple interface between cognitive and behavioral layer Cognitive Layer Stop … Activate … … behavior Override … Symbolic representations P1 P2 … Behavioral Layer B1 B2 … • … 27 Emotion expression using gestures Which emotion is expressed? 28 The End • I reached the end ;-) • Any additional questions comments suggestions ? 29 TODO • TeradaEtAl2008, A Cognitive Robot Architecture based on Tactile and Visual Information • Architectures don’t discuss plan repair, …? 30 GOAL Agent Programming Language GOAL agent program GOAL agent architecture See also: http://mmi.tudelft.nl/~koen/goal.html. April 9, 2015 31 DOD Levels of Autonomy http://www.fas.org/irp/program/collect/uav_roadmap2005.pdf 32 • • • • Tooth: http://www.kipr.org/robots/tooth.html Rocky III: http://www.kipr.org/robots/rocky.html Herbert: http://www.ai.mit.edu/projects/mobilerobots/veterans.html Robbie: http://www.magneticpie.com/LEGO/roverHistory/roverSize.html • B12 (Alfred): http://srufaculty.sru.edu/sam.thangiah/B12Robot.htm 33 Cognitive Architectures Overview Soar Scott D. Hanford, Oranuj Janrathitikarn, and Lyle N. Long, 2009, Control of Mobile Robots Using the Soar Cognitive Architecture 34 ACT-R 6.0 Architecture Current Goal ACT-R 6.0 Modify Declarative Memory Retrieve Check Test Pattern Matching And Production Selection Check State Motor Modules Schedule Action Identify Object Move Attention Perceptual Modules Environment 35 Cognitive Architectures Overview SS-RICS (2006) • SS-RICS = Symbolic and Subsymbolic Robotics Intelligence Control System • An extension of ACT-R • U.S. Army Research Laboratory, Aberdeen (Kelley and Avery) 36 Cognitive Architectures Overview ADAPT (2004) • ADAPT (Benjamin, Lyons, and Lonsdale 2004) Benjamin, P., Lyons, D., and Lonsdale, D., “Designing a Robot Cognitive Architecture with Concurrency and Active Perception,” Proceedings of the AAAI Fall Symposium on the Intersection of Cognitive Science and Robotics, October, 2004. 37