August 29, 2022
By Keith Shaw
Robot development continues to expand in many different areas, from warehouse robotics and industrial arms helping to create products and deliver goods through the supply chain; to agriculture robots and self-driving vehicles looking to provide more productivity, efficiency and safety for human workers and customers.
Other companies are working on developing a world where humanoid-style robots become a part of everyday life, with robots assisting in household chores and taking care of the elderly, or working in hazardous environments instead of humans. One problem, however, is that gaining access to the deeper levels of a robot’s capabilities, even to alter a simple movement or process, requires the expertise of highly skilled software developers and other experts.
Mollia, a Hungary-based artificial intelligence company founded in 2017, aims to transform the relationship between humanoid robots and humans from a “command-and-control” relationship to one of partnership. The company is looking to gamify robotic training and make it accessible to the general public. By using video games that run in physics-based simulation environments, the company is crowdsourcing user knowledge of locomotion in order to build a knowledge base of kinetic data. The motion data collected from these games can then be used to improve the kinematics of humanoid robots.
Robotics-World recently spoke with Daniel Vincz, Deputy CEO at Mollia, about the challenges and opportunities of humanoid robotics and their gamification approach. Vincz is a serial entrepreneur and computer engineer, having co-founded the INPUT Program, a program supporting Hungarian startups that was awarded the UN Global Best Practice award in 2018.
Robotics-World: What value do you see in humanoid robots being able to accomplish tasks that other robot designs (such as those that don’t look like a human) cannot accomplish? Why does the world need robots that look like humans?
Vincz: Out of all the things we know, humans possess the most advanced kinematics (we are bipedal, can use objects, easily learn complicated new tasks, and integrate complex motion cycles into other larger motion cycles). Therefore, utilizing this knowledge as a basis for developing robots is a natural approach, particularly involving humans in the teaching process directly. This approach requires a robot that has a similar physical build.
It is also important to note that humanoid robots have significant advantages when working in environments designed for humans. As a result, a humanoid robot may act as if it were a human in such an environment without the need to make any adjustments to the robot or the environment beforehand. There is also a psychological reason for using humanoid robots: Humans are more likely to relate to something that resembles them. In this way, we expect humans will have an easier time interacting with humanoid robots. The effect of these factors is to improve empathy, which also lowers the rate of vandalism. Furthermore, humanoid robots may require less training and manual reading since they are more intuitive. Additionally, they can be used as avatars to do physical work anywhere in the world from the comfort of a home or office. Humanoid robot research also benefits humans in other ways, such as robotic prosthetics, artificial intelligence, or crash tests, just to name a few.
R-W: What has been the main stumbling block in getting humanoid robots to move more like humans?
Vincz: There are several software-related stumbling blocks that humanoid robots have to overcome. For example:
- There is no way to describe human motion in a complete analytical manner, due to its incredibly complex nature.
- A robot trained by purely mimicking human motion will come with a few moves that resemble human movements. However, they will lack the “logic” behind these movements, and thus be unable to adapt them to different situations.
- Designing good interfaces so humans can improve a robot’s core abilities (rather than just teaching one particular move) is very challenging. In order to accomplish this, the robot’s AI needs to support specific hierarchical structures.
- Hardware is a serious limitation for physical robots as well, since human-like motions require very fine muscular control.
We believe that Mollia will be able to address most of the software-related blocks eventually, and by the time the hardware is good enough and affordable enough for mass production, we will have the software in place to enable ordinary people to collaborate intuitively with robots, regardless of their IT skills.
R-W: The key to having robots move more like humans appears to be within the AI software and the robot’s ability to make faster decisions based on perception. What advances has your company (or the industry in general) made over the past five years to improve AI software?
Vincz: In traditional AI applications, large amounts of data are collected and then fed into neural network-based systems that can aggregate this information without having to remember specific data points. This type of technology has been shown to be effective in a wide variety of applications, and has made rapid progress in recent years in areas such as image recognition and the creation of texts that are similar to those produced by humans. Nevertheless, it is worth noting that in most known solutions, it is difficult to gain a lot of insight into the internal workings of the trained AI or to improve its capabilities without providing it with additional information.
In our case, the kinematic kit of the robot is organized into a hierarchical system made up of elementary building blocks and rules that govern how they can be combined. This would be somewhat similar to a language-like grammar for robotic motion. By doing so, we are able to pinpoint entry points on the robot’s intelligence where input devices may be attached, allowing for intuitive interaction with humans and novel supervised learning processes. Users are able to transmit nuanced kinematic information to the system without having to convey it in formal terms. Over time, the demonstrated abilities will be integrated into the robot’s autonomous motion capabilities. In addition to the practical advantages of developing a robot’s motoric abilities, this also provides a high level of personalization.
It is important to note that while this does not eliminate the requirement for data collection, as the robot must acquire experience about the environment and its motoric functions, the collected data is organized into different structures, making it easier for developers and users to access.
R-W: What are some of the current hardware limitations that need to be improved in order to achieve the goal of more human-like motion from robots? Better gears, smoother motors, better sensors, etc.?
Vincz: For true human-like motion, a variety of hardware improvements are required, and the cost of many components must be significantly reduced in order for this technology to be commercially viable. Knowing the general trend of how rapidly these industries are improving, one has reasons to be optimistic. Furthermore, we believe that if we can provide software that can teach robots by leveraging the kinematic knowledge of humans, the resulting movements will be recognizable and relatable, even with relatively modest hardware.
R-W: What is the goal of creating a trainable AI NFT game? How will this improve the ability for robots to learn different skills or tasks?
Vincz: In the current development cycle, we utilize a physics-based simulation environment to provide an accurate representation of a realistic environment. In this way, we are able to involve a large number of users in the training process and assemble large sets of valuable data from their interactions with our system. In order to attract a large audience to participate in this co-development project, it is essential that it is both fun and entertaining. By using special video games, this can be accomplished.
With our technology, even children will be able to train humanoid robots of the future by playing video games. The players of this video game will be able to teach their virtual robot character a variety of kinetic skills and use these skills to compete against others. Considering our project focuses on community-based training, it is a natural fit for Web 3.0, blockchain, NFTs and play-and-earn games. This enables us to train robots at a large scale, as well as help people earn money in a fun and meaningful way.
The gamification of complex scientific problems has been a useful and successful approach in recent years; however, our literature review indicates that no attempts have been made in this area. We are not only interested in analyzing the collected data using machine learning methods, but also in allowing users to share robots and control techniques in order to improve the robot’s kinematic abilities collectively. As a result, one could say we are attempting to build a “Wikipedia for robotics motion.”
To learn more about the company and its video game development for robotics kinematics, visit the Mollia website here.