TOKYO -- Fujitsu Laboratories Ltd. has said that it has developed a learning system for humanoid robots that uses a dynamically reconfigurable neural network to enable the efficient learning of movement and motor coordination.
The company said the neural network approach makes the teaching of motion to a humanoid robot simpler and faster. Traditional approaches have required large amounts of software to be written to connect and control a robot's complex sensors and actuators.
Neural networks have so far proven to be very slow in learning movements, requiring days or even months, and they have been unable to generate a wide variety of motions, according to Fujitsu.
Fujitsu Lab's technology is based on an architecture of Central Pattern Generator (CPG) networks, mimicking a function found in earthworms and lampreys, which mathematically simulates a neural oscillator, which mimics a function found in vertebrates providing a neurological 'pulse'.
This advanced neural network features have been combined with a numerical perturbation (NP) method that quantifies the configuration and connection-weight status of the network.
This combination, known as CPG/NP learning, has been optimized and applied to motion learning in the new technology. In addition, Fujitsu simultaneously developed a software program, known as the Humanoid Movement-Generation System, which enables humanoid robots to learn a wide range of movements.
The central pattern generators generate motions in the robot using self-induced oscillations. These are evaluated using a pre-set evaluation function that determines whether the movement is correct or incorrect.
A big advantage is that the approach greatly minimizes the amount of software code involved in motion control, cutting it to less than one-tenth that used in conventional systems, Fujitsu said. For a robot with 20 moveable joints the learning time is reduced by 10 to the -30th power of the time previously required, Fujistus said. This enables the robot to learn to adapt and react instantly as it moves.
As part of the robot-motivated development work Fujitsu has also developed a programming language for coding neural networks. Later this year the software system will be included in the HOAP Series of humanoid robots that are already being marketed by Fujitsu Automation Limited
Fujitsu said more details of the system would be made public at Robodex 2003, which scheduled to take place at Pacifico, Yokohama, Japan from April 3 to 6.
