For many people, the thought of robots might call to mind their use in product assembly in large factories or the sorting of packages in logistics hubs. Currently, robotics researchers around the world are focused on developing robots that work alongside humans and provide support. However, working alongside humans means that robots need to enter various locations with people and provide reliable support for tasks in diverse environments. A particularly crucial challenge is assisting or replacing staff in workplaces faced with labor shortages and retiring workers.
To meet this challenge, Hiroshi Ito, a senior researcher at Hitachi's Research & Development Group and a guest junior researcher at Waseda University's Future Robotics Organization, has developed a deep learning algorithm called 'deep predictive learning' and applied it to robots. This technology is based on the predictive coding principle, an information processing mechanism in the brain, and dramatically reduces the extensive trial-and-error and learning that had been a major obstacle in training robots to perform intended movements.
A robot with deep predictive learning can master complex movements and whole-body coordinated actions simply through learning for a few hours on a computer after a human demonstrates the target movements multiple times (typically around 10-15 times) by operating the robot remotely. This allows end users without programming expertise to easily train robots to perform desired actions. Notably, this method enables robots to acquire tacit knowledge and skills by learning from the movements of skilled workers, which was difficult to achieve with conventional robotics technology.
The deep predictive learning developed by Ito is now being incorporated by some companies into their robot development. A robot developed jointly by an AI startup company and a domestic space agency for opening and closing soft zippers in outer space was created with reference to a paper co-authored by Ito. Ito has released an open source library implementing a deep predictive learning algorithm called EIPL (Embodied Intelligence with Deep Predictive Learning), aimed for global adoption of the technology.
Ito has also developed a research robot that can easily be taught complex movements. The robot is designed so that an operator can train it to perform a variety of complex movements by reaching their arms around from behind the robot, allowing the robot to collect high-quality learning data from the same perspective as the operator. The high-quality data collected in this way will be used to train various deep-learning models. The ultimate goal, Ito says, is to create a robot that can independently contemplate and perform movements it has not been explicitly taught.
