Photo credit: www.sciencedaily.com
Researchers from Cornell University have introduced an innovative robotic framework enhanced by artificial intelligence, referred to as RHyME (Retrieval for Hybrid Imitation under Mismatched Execution). This new system enables robots to acquire skills by observing a single instructional video.
Traditionally, training robots has been a tedious process, requiring meticulous, detailed instructions to help them complete even the simplest tasks. These machines often struggle when faced with unexpected situations, such as dropping an object or misplacing a tool. According to the research team, RHyME promises to streamline the development and implementation of robotic systems by significantly decreasing the amount of time, resources, and financial investment needed for training.
Kushal Kedia, a doctoral student specializing in computer science, noted, “One of the frustrating aspects of working with robots is the necessity of gathering extensive data for various tasks. Humans, on the other hand, learn by observing others.”
Kedia is set to present the research paper titled “One-Shot Imitation under Mismatched Execution” in May at the Institute of Electrical and Electronics Engineers’ International Conference on Robotics and Automation, taking place in Atlanta.
The vision of home robot assistants remains distant, largely due to the current limitations of robots in navigating the complexities of the physical world. To accelerate their learning curve, researchers like Kedia are utilizing instructional videos that showcase how humans perform different tasks in controlled settings. This technique, known as “imitation learning,” aims to help robots quickly grasp task sequences and adapt to varied real-world scenarios.
Despite this innovative approach, there are underlying challenges. Human movements are often too fluid for robots to replicate precisely, and training with video requires extensive footage. Additionally, demonstrations must be executed slowly and perfectly; any discrepancies between the actions performed by the human and the movements of the robot have historically hindered effective learning.
RHyME represents a solution to this challenge by offering a scalable framework that enhances a robot’s adaptability. This system empowers robots to leverage their own memories to perform tasks after viewing them only once while utilizing inspiration from similar actions in other footage. For instance, a robot powered by RHyME that sees a human retrieve a mug from a counter and place it in a sink will search its database of videos to find comparable actions, such as grasping an object or manipulating a utensil.
With RHyME, robots are now capable of mastering multi-step tasks with a drastically reduced amount of data needed for training—only 30 minutes of direct robot activity is required. In experimental settings, robots trained using this new framework demonstrated over a 50% increase in success rates for completing tasks compared to traditional training methods, according to the researchers.
Source
www.sciencedaily.com