Abstract:In robotics community, inspiration from human development theory offers a promising way for robots to efficiently achieve many abilities. Humanoid robot reaching, as an essential ability which forms the foundation of grasping and manipulation, is also possible to take advantages from human. For human infants, reaching is emerging around 3–5 months of age, which is also a fundamental skill for the development and refinement of future high level motion or cognitive behaviors. The issue of how infant reaching is developed has been investigated for decades and several views were established and discussed. Recently, Corbetta et al. proposed that the emergence of reaching is the product of a deeply embodied process, in which infants first learn how to direct their movement in space using proprioceptive and haptic feedback and then map the visual attention onto these bodily centered experiences. With this new sight, this paper addresses the problem of how a robot develops its reaching ability autonomously, and a novel infant-inspired model is proposed. The model is composed of several blocks so that to closely capture the inherent mechanisms of infants in the process of developing the reaching skill. To evaluate the proposed model, a child-sized physical humanoid robot PKU-HR6.0 is employed. Just like an early infant, the robot is assumed without any other abilities. It is only equipped with the proposed model with random initialized parameters. The reaching ability is expected to be developed all by the robot itself. Through iteratively babbling arm motions in its workspace, the robot firstly acquire a reliable sense of its body and movement in space, and then map the sense of the vision onto the proprioception when contingencies happen. Experimental results show that the robot equipped with the proposed model achieves the reaching ability effectively and successfully in a completely autonomous style. The robot is able to reach the objects in different real environments with a high performance, and can even grasp the objects with an average successful rate of 82.50%. Our experimental results also verify that the movement of one joint may induce another, and some sepecific joint may play dominant role in certain category of movements.

Learning and Acting in Peripersonal Space: Moving, Reaching, and Grasping

Dexterous Manoeuvre Through Touch in a Cluttered Scene

Active exploration for body model learning through self-touch on a humanoid robot with artificial skin

An Infant-Inspired Model for Robot Developing Its Reaching Ability

Perception-enhancement Based Task Learning and Action Scheduling for Robotic Limb in CPS Environment

Neurally Plausible Model of Robot Reaching Inspired by Infant Motor Babbling

Experiments with a Social Learning Model

Learning active tactile perception through belief-space control

Learning body models: from humans to humanoids

Learning at the Ends: From Hand to Tool Affordances in Humanoid Robots

More Than a Feeling: Learning to Grasp and Regrasp using Vision and Touch

A human-like learning approach to developmental robotic reaching

Learning Robotic Manipulation through Visual Planning and Acting

How Does a Robot Develop Its Reaching Ability Like Human Infants Do?

Active Perception and Representation for Robotic Manipulation

A Survey of Embodied Learning for Object-Centric Robotic Manipulation

Active inference body perception and action for humanoid robots

A Grasp Pose is All You Need: Learning Multi-fingered Grasping with Deep Reinforcement Learning from Vision and Touch

Learning Models For Constraint-Based Motion Parameterization From Interactive Physics-Based Simulation

Self-developing Proprioception-Based Robot Internal Models

Stage-Wise Learning of Reaching Using Little Prior Knowledge