Thu. Sep 28th, 2023
    Robots Can Learn New Skills Overnight, Says Toyota Research Institute

    The Toyota Research Institute (TRI) has made advancements in teaching robots new skills overnight, according to TRI CEO and Chief Scientist Gill Pratt. Traditional machine learning required millions of training cases, which was time-consuming and impractical for physical robots. However, TRI’s system uses more efficient robot learning techniques and diffusion models to teach robots. TRI has successfully trained robots on over 60 skills using this method. The system allows robots to learn skills that can be applied in diverse settings, making them adaptable to different environments. This is crucial for robots to function effectively in real-world scenarios, as they can encounter unstructured and changing environments.

    TRI’s focus on developing robots that can help older people live independently is particularly important in places like Japan with an aging population. The goal is to create a system that can operate in different environments and adapt to changes. Traditionally, roboticists have had to anticipate and program robots to manage various edge cases and deviations. Teleoperation, where a person remotely drives a robot through demonstrations, is commonly used in robot learning. TRI’s system utilizes teleoperation along with force feedback to improve training. The person controlling the robot can feel what the robot is doing, enabling more coordinated interactions.

    The system collects data from sight and force feedback to produce a comprehensive understanding of the task. Tactile sensing plays a crucial role in replicating dexterous behaviors accurately. TRI’s experiments with tactility have shown promising results, such as a 90% success rate in flipping pancakes. Once the training aspect is completed, the neural networks continue to train overnight. The diffusion policy used by the system allows robots to learn behaviors by representing them as denoising diffusion processes.

    Overall, TRI’s advancements in robot learning highlight the potential for robots to acquire new skills quickly and adapt to different environments. This progress is a step towards developing more flexible and capable robots for real-world applications.

    – TechCrunch:
    – Toyota Research Institute (TRI)