Startpage privacy
       






MRI brain

Project Overview

 
  • Autonomous robots are smart robots that are capable of making decisions to interact with the environment in exploration and learn from the experience.

Rationale

 
  • Autonomous robots can mimick the basic functions of simple organisms, such as insects, such that the basic principles of operation used to interact with the environment can be explored.
  • In designing an autonomous robot, we need to address the issues of how an organism, such as an insect, solves the problems encountered in the interaction with the environment.
  • That is to say, the organism needs to "know" or "acquire knowledge" about the environment so that it can respond appropriately in the changing environment in order to survive.
  • Thus, an autonomous robot faces the same issues as an insect would have encountered.

Advantages of Using Autonomous Robots as a Model

 
  • The reason why autonomous robot is a good model for investigating the evolution of the nervous system is that we can explore various principles of operation without being constrained by the inherited evolutionary path.
  • The evolutionary paths of insects are often stuck due to its heritary past, whereas we can explore many more scenarios in neuroengineering.

Real Robots vs. Simulation

 
  • Real robots need to deal with the unexpected environments, such as wheels getting stuck. These real-world issues pose a set of challenging problems for the organism to solve.
  • Simulations often assume that the robot will move without spinning the wheels, but that is not a real-world problem. Insects need to solve the same problem when they are stuck while they encounter obstacles.

Research Objectives

 
  • Find the simplest/minimal set of solutions to solve the most/complex problems – this is called finding an "elegant solution."
  • That is to ask: What is the simplest nervous system an animal can have in order to solve the most complex problems encountered by a simple organism, such as an insect.
  • We choose not to study complex high-level vision, such as scene analysis, because it is an overkill to the problem.
  • Instead, we choose to study simple compound eye without even forming a retinal image, yet insects can detect a great variety of visual objects, and escape from predation without even having a brain!
  • Autonomous robots are similar in using the simplest set of algorithms to solve the most complex problems without relying on a super-computer to solve these problems.

Basic Components of Autonomous Robots

 
  • Sensory inputs (sensors for detecting light, touch, sound, etc.)
  • Motor outputs (actuators for producing motions, such as wheeled robots or legged robots)
  • Controller (integrators of inputs to produce outputs, including memory modules to store "experience")

The Challenge

 
  • Design a robot without any a priori knowledge of the environment.
  • That is, the robot has no knowledge of the external environment, and there is no pre-programming done to allow the robot to know beforehand what to expect or how to respond in a given solution.
  • The robot has to learn from the exploratory experience to acquire knowledge about how to respond appropriately in the environment in order to survive.
  • If we pre-program the solution of how to solve the problem, we would be actually cheating, and defect the whole purpose of studying the robot's behaviors.
  • So the question becomes: What is the minimal set of assumptions and constraints that we need in order for the robot to acquire knowledge from the exploration in the environment such that the internal circuitry will be self-organized to produce the final appropriate response similar to those we found in the reflexes in insects.

The Solutions

 
  • Neural network algorithm is used to "learn" from experience. Exploration by the robot provides the exemplary set of data to build an internal map of the environment.
  • Neural nets are known to adapt to the environment, and learn from the environment without any pre-programming. The task of the neural net essentially construct the nonlinear mathematical mapping functions from the input set (external environment) to the output set (motor responses), with the internal synaptic weights stored as the internal map.