The biological basis of hitting or kicking the ball can be better understood by copying the mechanisms of learning in the cerebellum. Extracting information from sensory motor signals, this leaf-shaped organ responds to environmental cues. The BIOMODULAR project built a novel bio-inspired computational learning model for modular robots. “Merging machine learning techniques and a spiking modular cerebellum, we developed a process that leads to the formation of long-term motor memories, a cerebellar-machine learning (CML) network,” outlines Dr Silvia Tolu, project coordinator and investigator. Hitting that ball better and better Dr Tolu explains further what BIOMODULAR achieved: “We didn’t aim to build high-performance robotics directly, but rather systems that are able to adapt and learn from actual self-training, such as manipulating an object towards improving their performance.” The resulting robots are very flexible for a wide variety of tasks and scenarios. Furthermore, the control engine can adapt to a robot and optimise its performance at each stage. Researchers built bio-inspired control loops that embed a CML network and tested them with both simulated and real robots, as well as the neuromorphic hardware SpiNNaker, performing tasks with changing kinematics and dynamics. Along with success come challenges and the SpiNNaker board had its drawbacks. Currently, it is not possible to code different plasticity rules involving the same neurons. “This is something still under research and development and once this is possible, the spiking cerebellar network with three plasticities rules will be implemented on the chip,” Dr Tolu emphasises. The next generation of robots BIOMODULAR has promoted the development of artificial adaptive learning systems embedding spiking neural networks and machine learning mechanisms. With wide applicability in robotics, better understanding of the brain will aid in the design of plausible biological control schemes that can be generalised to any robot, under any conditions. “As a matter of fact, mimicking the biological functionality of the central nervous system will lead to creation of autonomous intelligent robot agents as the next generation of robots,” Dr Tolu points out. Not as extreme but reminiscent of Asimov’s robot literature and ‘I, robot’, robots of the future will perform in real environments, maybe partially unknown and/or changing, as living systems do, under control paradigms that go beyond the scope of conventional control algorithms as they self-adapt, learn internally and self-recognise. Robots will operate safely in proximity with people and in different fields away from the tightly controlled environment of the factory floor. The future generation of self-learning and compliant robots can operate safely in a human environment. Transforming society, these would have massive impact on technology as autonomous systems for assistive robotics. Back to here and now The BIOMODULAR approach reduces the amount of information the control system needs, leading to a real-time robotic control system that can learn autonomously how to perform determinate physical tasks and adapt to changing, sometimes harsh conditions. “Excessive customisation will not be necessary,” stresses Dr Tolu. Any kind of robots will benefit from this adaptive predictive control system for performing desired, task-fulfilling behaviours. “Exploiting this approach, the Marie Curie-funded BIOMODULAR continues to pursue the discovery of important insights into the modular structure of the cerebellum, and its involvement in processing sensory input for motor control tasks,” Dr Tolu concludes.
BIOMODULAR, robot, cerebellum, modular, machine learning, control system, sensory, CML network