developmental skill aquisition with general purpose actuator

an investigation into methods of developmental skill aquisition with general purpose actuator

• to identify existing developmental learning methods that is capable of developmental learning process with a finite, pre-defined set of actions.
• to experiment aforementioned methods and evaluate efficiency and effectiveness of each
• to find or develop developmenatl learning method that allows robot/agent to acquire skill automatically without physical or cognitive(software) modification.

Developmental Learning/Developmental Robotics is a machine learning method with focus on automatic knowledge development.¹ The core tension of the method is to let agent autonomously construct knowledge and self-programme, base on experiences/experiments of their interactions with the world. The method take inspiration from children’s cognitive development. inapp On the various definitions and properties of Artificial General Intelligence, one is the ability to develop, autonomously, knowledge/internal representations that is effectively incremental², meaning that the agent should be able to acquire new knowledge and new skills of new subject domain all by itself, with as little friction across boundaries of disciplines as possible. Developmental Learning is one paradigm of such.

There are mainly 2 lines of works on this subject, one focuses on modelling infant’s cognitive development³ and the other focuses on development of agent with analogies/features from cognitive development and approaches from computer science and machine learning⁴.

On the infant line, visual signal processing, eye-gaze tracking are normally present, while on the robotics line, general purpose actuator like arm and wheel, and skill acquisition with those actuators are nomrally present.

This work would aim at evaluating and developing developmental learning schemes and their application on general purpose actuators such as robotic arm, in pursuit of a learning scheme that is general-applicable(architecture-free) and can let agent effectively acquire new skill with a pre-defined, finite set of actuators or actions.

• Handelling of signal in abstract(logic; machine learning) and in raw(signals processing; embodied robotics)
• Literature review, experiment and evaluation/organization on existing approaches
• Environemnt and task/skill design.
• Review and develop knowledge representation and self-programming techniques.
• primitive action engineering: complex action made from primitive actions.