Empowering Robots with Intelligence: Innovations in Industrial and Social Robotics

Our research revolves around the development of intelligent robots and autonomous systems. We focus on two key domains: industrial robotics and social robotics. In these domains, we encounter several challenges that motivate our work. 

In the industrial sector, robotic systems have been successfully integrated into various manufacturing segments, offering economic benefits in terms of performance, scale, and production quality. The robots excel in automating high-volume production of standardized products. However, as the industry shifts towards mass customization with shorter product life cycles, smaller batch sizes, and increased product variations, the limitations of current industrial robots become apparent. These robots lack flexibility to adapt to changes in part specifications, locations, sizes, and shapes. They require time-consuming reconfiguration by expert engineers, and even motion planning methods still rely on high-level programming by specifying goal locations and spatial way points. Thus, there is a need for more flexible and adaptable industrial robots that can address these challenges.

In the realm of social robotics, the growing aging population and the shortage of healthcare workers present significant concerns. Caring for the elderly and improving their quality of life has become a pressing issue. Socially-assistive robots offer a potential solution by providing companionship to older individuals. We believe that enabling socially-assistive robots to interact with their environment can serve as surrogates for human assistants, especially for elderly people with physical impairments.

Both domains share a common requirement: robots need to coexist with humans while attending to tasks commanded by humans. Instead of solely focusing on fully autonomous solutions, which are incredibly challenging, we have embraced the concept of semi-autonomous methods. We have embarked on exploring a framework of human-robot symbiosis, where humans and robots form a unified entity, enabling them to accomplish tasks that neither can perform alone. Our framework, known as the “Internet of Skills” (IoS), aims to equip robots with advanced skills learned from a limited set of expert demonstrations. These skills can be packaged as software and integrated into other robots with similar capabilities. Teaching robots new skills remains a challenge, but we have leveraged human demonstrations to address some of these issues. By directly observing expert actions, the robot can learn the goals of demonstrated skills. When unsure how to act, the robot can refer to a cache of expert experience and mimic expert’s actions. By incorporating expert demonstrations into the training process, we significantly reduce the time it takes for robots to learn new skills and perform tasks that were previously too complex to be done autonomously. This framework establishes a tight coupling between human and robot actions, allowing each agent to compensate for the other ‘s deficiencies. We acknowledge that not all tasks can be fully automated with current methods and that some will require human intervention. Through this semi-autonomous system, robots can learn incrementally from human demonstrations, expanding the range of tasks that can be automated over time.

We have successfully trained robot agents using expert demonstrations to complete a variety of tasks and integrated these demonstrations into our learning framework (Figure 1). 

Figure 2: Demonstration of a Kinova Gen3 7DOF robot following the expert’s joint angles.

To optimize the policies learned through demonstration, we have employed reinforcement learning, enabling the student agents to potentially surpass the performance of the expert demonstrators. This technique offers benefits such as achieving higher precision than humans and filtering out hand shaking during delicate manipulation tasks. Moreover, we recently extended our training framework to both a Kinova GEN-3 robot  (Figure 2) and a humanoid robot (Nao), (Figure 3) despite the significant references in arm configuration. This exemplifies the versatility and power of our framework.

Figure 3: Nao grasping task

The human/robot coupling in our framework serves as a bridge between tasks that are too complex for complete robot automation yet not feasible for humans to perform. For instance, partial automation can enhance the capabilities of robots operating in hazardous environments or allow human technicians to take control after a failed autonomous inspection. This operational mode is particularly important in highly regulated industries like healthcare, where a specialized expert must maintain full control over the robot at all times. 

In conclusion, the development of robots and autonomous systems that embody intelligence has the potential to revolutionize both industrial and social robotics. The Internet of Skills framework offers a promising approach to bridge the gap between human and robot capabilities, fostering a symbiotic relationship where each entity compensates for the other's limitations. By leveraging expert demonstrations and reinforcement learning, robots can acquire advanced skills and optimize their performance. This framework paves the way for a future where robots seamlessly integrate into our society, enhancing productivity in industries and improving the quality of life for individuals in need of assistance. With ongoing research and technological advancements, the vision of human-robot symbiosis is steadily becoming a reality, shaping a world where humans and robots coexist harmoniously, achieving tasks that were once unimaginable.