Affordance Learning Methods In Developmental Robotics

The goal of developmental robotics is to endow the robots with biological capabilities such as perception,behavior,learning and decision,so that they could help people finish different tasks in dynamic environment.To learn from the nature is the fundamental way to carry out this research,and to produce an infant-like robot is the first target.In 1977,the American psychologist J.J.Gibson proposed the Affordance theory.He stated that an infant first perceived the affordances of the environment and then perceived its attributes.Since the 21 st century,the affordance theory was introduced into developmental robotics,where it encodes the relationship between the robot and environment in terms of an action that the robot is able to perform.This potential action is jointly determined by the current task,the robot's capabilities and the environment's essential properties.The notable character is that an object is perceived by its affordances(e.g.,graspable,moveable),instead of identities(e.g.,color,shape).The focus of affordance research is to learn the potential action relationships between the robot and environment.A robot uses these affordances to finish tasks during which its perceptive,behavioral and learning capabilities are developed continuously and incrementally.This dissertation first introduces the concept of infant learning and intrinsic motivation,then analyzes three research fields in developmental robotics such as autonomous mental development,embodied intelligence and affordance.We classify the affordance learning methods of an infant before 2-year-old into five stages.The first three stages are associated with simple tasks,while the interations in the later stages are more and more complex.The survey of affordance research in developmental robotics is presented in Chapter 2.This dissertation is mainly to model the five stages of affordance learning,and its contributions are as follows:(1)Define the affordance from the perspective of developmental robotics,and design three incremental affordance models as well as the corresponding learning methods for simple tasks.The "discrete state space to action" model uses intrinsic motivation functions to describe the reward of affordance relationships.The "continuous state space to action" model fits the mapping from continuous state space to actions.The "visual-speech inference" model abstractly represents the visual and behavioral information through speech,and uses tabular relations to infer the unknown information.As a result,this model reduces the number of parameters during human-robot interaction,and promotes the learning mechanism from single-dimension to two-dimension.(2)Propose the affordance prediction methodology and its formalism at the subtask level.Reduce the state space by decomposing the whole task into hierarchical task graph and introducing state abstraction machanism.Predict the dynamic affordances based on the subtask strategy and goal-free affordances.Promote the influence of an affordance from a reactive action to a subtask's strategy.Experimental results indicate that this dynamic affordance-based approach performs better than the static affordance-based approach.(3)Propose the affordance matching approach based on shared information in multi-robot.This approach describes a robot in terms of its functional capabilities,and uses a sharing mechanism to provide information for all the robots.As a result,a robot could match the affordance relationships without necessarily physical interaction.Transportation experiment indicates that the efficiency of our approach is higher than the non-affordance approach.(4)Develop a robotic simulator software,CogRSim.CogRSim is used for affordance research,and its main modules are affordance detection and execution.The advantages of CogRSim include:(i)3-dimensional scene could be viewed synchronously in 2-dimensional panel;(ii)different objects could be assembled into an integrated and complex object,which has arbitrary shape and unified physical attribute.CogRSim's usability is verified in both single-robot and multi-robot experiments.