For some years now, thanks to the incessant work of many researchers, advances in artificial intelligence and machine learning have made it possible to have better robotics today. For example, a few days ago we learned that scientists had managed to develop a first robot capable of returning the smile (and many other facial expressions) to those around it.
But despite this, there is still a huge gap between what humans can do and what robots can actually do. Precisely, it is the objective of many researchers to try to overcome this series of obstacles still existing in the manipulation of robots. Or, more specifically, in their ability to manipulate environments, use objects and adapt to changing stimuli .
As many experts say, in robot manipulation, learning becomes a promising alternative, proving great success, especially in pick-and-place tasks.
Recently, in an article published in Science Robotics , experts from the Department of Computer Science and Engineering at Lehigh University summarize, compare and contrast the existing research so far on robot manipulation through the lens of adaptability. , emphasizing the usefulness of modularity in the design of machine learning, and also point out the need to carry out appropriate representations for manipulation tasks.
As they indicate, “the proper modularization of learned manipulation skills can open ‘black boxes’ and make them more explainable.” And what are the so-called “black box” solutions? It means that researchers may not know when and why a certain learned skill fails.
In order to find possible solutions, the experts propose in their article a total of eight particularly promising areas in which it would be necessary to advance to achieve a better capacity and adaptability of the manipulation of learned robots:
- Representation learning with more sensory modalities, such as tactile, auditory, and temperature cues.
- Advanced manipulation simulators to make them as realistic and fast as possible.
- Customization of tasks and skills.
- “Portable” task representations.
- Informed exploration for manipulation in which active learning methods can efficiently find new skills by exploiting contact information.
- Continuous exploration.
- Massively distributed active learning.
- Hardware innovations that simplify more challenging manipulations.
Following some of these guidelines, researchers are currently working on tactile sensorimotor skills , with the goal of making robots more dexterous and robust. Especially at a time when current research related to the manipulation of learned robots is still in its infancy.
And they coincide in pointing out something very important: “the promising future and the vast space for exploration will make the manipulation of learned robots an exciting area of research for the next decades.”
And, although there are still many questions related to research to be answered, there is no doubt that improving this area could lead to the development of domestic robots. “We may see robots cleaning our tables or organizing cabinets in the near future.”
