Affordance-based Perception, Learning and Planning using Range Images Erol Ş ahin KOVAN Research Lab. Dept of Computer Eng. Middle East Technical University Ankara, TURKEY http://kovan.ceng.metu.edu.tr Dept. of Computer Engineering 1 Middle East Technical University
Developmental/Epigenetic Robotics Behavior development in psychology J. Piaget & E.J. Gibson Controlled Discovery of Repeated Reflexes action capabilities interaction Affordances Use relations Discover Interact Simple in goal-directed general with the pre-coded behavior relations environment behaviors M.Cakmak, M.R.Dogar, E. Ugur and E.Sahin. Affordances as a Framework for Robot Control . Proc. of 2 EpiRob’07.
Learning of Affordances • Differentiation: discovering distinctive features and invariant properties in the environment • Exploratory activities that bring about changes in the environment that an action produces • Perceptual Learning: develop an anticipation of outcomes based on perception of invariants, actions become performatory E.J. Gibson (1910–2002) 3
Equivalence Classes Behavior Equivalence Entity Equivalence (effect, (entity, <behavior>)) (effect, (<entity>, behavior)) Effect Equivalence Affordance Equivalence (effect, <(entity, behavior)>) 4
Experimental Framework (1/2) • 6 wheel, differential drive • MACSim : High-fidelity simulation environment • SICK laser range finder • ODE used as physics engine • 3-D scanning • Sensors and actuators are – 0.25° horizontal resolution calibrated (180°) – 0.23° vertical resolution (165°) 5
Experimental Framework (2/2) Primitive behaviors Limited motor capability Pre-coded 6
Perceptual Features More than 30000 perceptual features! 7
Part 1: Perceptual Learning of Affordances 8
Traversability for KURT3D 9
Representation of entity and effect A single interaction Perceive Act Get result Features(initial) MOVE_FORWARD Success/Fail 3000 x entity behavior effect • E.Ugur, M.R.Dogar, M.Cakmak and E.Sahin. The Learning and Use of Traversability Affordances using Range Images on a mobile robot. ICRA, Rome, Italy, April 2007. 10
Learning: Selecting relevant features 3000 interactions Effect: Filtered Effect: Entity success Entity success Filtered Effect: Effect: Entity Entity fail fail Feature . . Selection . . (ReliefF) . . Effect: Filtered Effect: Entity fail Entity fail 11
Perceptual Economy Scan only this region Only 1% of the features are relevant! 12
Learning: Mapping entity to effect 3000 interactions Target Effect: Filtered success Entity values Effect: Filtered fail Entity . SVM . input . Effect: Filtered fail Entity 13
Prediction of Traversibility for Novel Objects TRAIN TEST prediction prediction 100% 86% 100% 100% 83.8% Successful generalization over novel objects 14
Traversability in a cluttered environment 15
Traversibility on KURT3D 16
Pass-thru-ability on KURT3D • The training used only single object interactions. • The robot has not concept of an object or a gap. • The robot does not have any idea on the size of its body. • Yet, one can see an affordance ratio here.. How would Warren et al. comment? 17
Curiosity-Driven Learning • Large number of training samples were required: – ~3000 virtual interactions with environment. – Learning process is costly, time-consuming, risky. • Minimize number of interactions with minimal degradation in learning process. 2 phase learning: – Bootstrapping: small number of interactions • Learn the relevant features • Initiate an SVM model. – Curiosity-driven • Interact with the environment and update SVM only if the current situation is an interesting one • E.Ugur, M.R.Dogar, M.Cakmak and E.Sahin. Curiosity-driven Learning of Traversability Affordance on a 18 Mobile Robot . ICDL, London, UK, July 2007.
Not interesting! Probably traversable Probably non-traversable No object in the vicinity Cylinder object is very close. 19
Interesting Object is located Cylinder’s surface is at the boundaries for similar to sphere’s Go-forward action 20
Curiosity-Driven Learning PERCEPT SVM (trained upto now) • Within curiosity band • Outside curiosity band • Classifier is less certain about hypothesized effect • Classifier is more certain about hypothesized effect • Execute behavior • Do not execute behavior • Learn from this sample. (Update SVM). • Skip this sample. 21
Two parameters of learning • Bootstrapping duration: No improvement after a certain level (25) • Width of curiosity-band is optimized. (0.5) • With bootstrapping of 25 and curiosity of 0.5, 200 interactions delivers as good/better performance than 3000 interactions! 22
Performance comparison Curiosity-based interaction is more economical 23
Part 2: From Primitive to Goal-Directed Behaviors 24
Representation of entity and effect A single interaction Perceive Act Perceive Features(initial) MOVE_FORWARD Features(final) Features(final-initial) 3000 x entity behavior effect M.R.Dogar, M. Cakmak, E. Ugur and E. Sahin. From Primitive Behaviors to Goal-Directed Behavior Using Affordances. IROS, San Diego, October 2007. 25
Learning: Cluster Effects For each behavior: Effect Effect mean 3000 interactions category-1 prototype-1 Entity Effect Entity Effect mean Effect K-means Effect . prototype-2 category-2 (k=10) . . . . . . . Entity Effect Effect mean Effect category-10 prototype-10 No fail/success criteria! 26
Learning: Selecting relevant features 3000 interactions Effect Filtered Effect Entity category Entity category Filtered Effect Effect Entity Entity category category Feature . . Selection . . (ReliefF) . . Effect Filtered Effect Entity category Entity category 27
Learning: Mapping entity to effect Target 3000 interactions Effect Filtered category Entity values Effect Filtered category Entity . SVM . input . Effect Filtered category Entity 28
Goal-directed Behaviors Behavior … Behavior Perception Execution Perception Selection Selection 29
Goal: Approach 30
Goal-directed Behaviors on KURT3D Goal is provided at run-time not during learning! 31
Blending Behaviors Motor parameters 0.4 x Weighted sum Similarity 0.6 x 0 x Inspiration source: Population coding in motor cortex • M.R.Dogar. Using Learned Affordances for Robotic Behavior Development. M.Sc. Thesis, Middle East Technical University, Ankara, September 2007. • M.R.Dogar, E.Ugur, E.Sahin and M.Cakmak. Using Learned Affordances for Robotic Behavior 32 Development. . Accepted to ICRA’08.
Goal: Approach Using primitive behaviors Using behavioral generalization Primitive behavior Blending of behaviors 33
Blending Behaviors on KURT3D Behavior blending allows to span a whole range of behaviors from a limited pre-coded primitive behaviors 34
Part 3: Planning with Learned Affordances 35
Behaviors Movement Behaviors Lift Behavior • E.Ugur, M.R.Dogar and E.Sahin. Planning with Learned Object Affordances . Submitted to AAAI’08. 36
Perception 37
Predicting the next state of the entity • Adding the effect prototype of the behavior to be applied to the current entity representation provides us a prediction of the expected state of the entity. 38
Plan Generation 39
Sample Plans 40
Analysis of effect classes 41
Learning Process 42
Planning Performance 43
Goal: Activate the Button • Goal: – Any object: The range image coordinate features should be high – Button object: Mean distance should be small 44
Planning on KURT3D 45
Acknowledgements Maya Cakmak Mehmet R. Dogar Emre Ugur Multi-sensory Autonomous Cognitive Systems supported within the Cognitive Systems Call of FP6-IST (FP6-IST-2-004381) 46
Recommend
More recommend