Baby Gym For Robots: A New Platform For Testing Developmental Learning Algorithms

This extended abstract describes a new platform for robotic manipulation research that was inspired by some of the first toys that human infants learn to manipulate. It summarizes the results of our existing research on pressing buttons and formulates some ideas for future work. Motivation and Inspiration Figure 1: My son in his Baby Gym. When my son was growing up I noticed that it took him almost 5 months to learn how to press a button reliably on one of his toys. The toy had a number of buttons, sliders, levers, and knobs that could be manipulated (see Figure 1). These toys come in many varieties and in recent years some toy manufacturers have started to call them Baby Gyms. Initially my son was more interested in exploring the plastic casing of the button instead of the button itself. He had to learn what the button looks like and how and where to press it. This learning process started when he was 4.5 months old and continued until he was 9 months old. Once he had mastered this skill, however, he knew how to detect other buttons around the house. Furthermore, he knew that these devices are for pushing and he knew how to push them. Light switches, keyboard keys and microwave buttons were now intuitively obvious to him as they were in the same equivalence class with that first button. Similar developmental sequences were observed for a number of other devices present in the baby gym. Copyright © 2011, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved. Learning to Press Buttons The devices found in baby gyms – buttons, sliders, levers, wheels, and knobs – are widely used in human environments. Buttons alone are present in virtually every gadget that humans have ever created. Robots operating in human environments would have to know not only how to navigate without hitting any obstacles, but also how to operate these devices by actively touching and manipulating them. Otherwise these robots would not be very useful. These manipulation tasks are still challenging for robots yet one-year-old infants can perform them quite easily. This suggests that the exploration methods that infants use may hold the key to solving such tasks with robots. These methods rely on exploratory behaviors such as pushing, scratching, and nudging, which can be quite powerful when combined with multimodal perceptual change-detection functions. One key advantage of the baby gym platform is that it offers a nice way to have multiple reproducible experiments. These observations motivated a pilot study to test if a robot can learn to press buttons in a similar way (Sukhoy et al. 2010; Sukhoy and Stoytchev 2010). That is, to test if a robot can learn both where and how to press buttons from its own experience without prior knowledge of what buttons look like. We built an experimental fixture similar to a baby gym (see Figure 2) that contained multiple doorbell buttons. The results showed that the robot was indeed able to learn to press buttons autonomously. Furthermore, the robot simultaneously learned a visual model for what a button looks like and was able to use it to detect novel buttons. Figure 2: The experimental fixture with doorbell buttons inspired by my son’s baby gym. Figure 3: Mapping between the baby gym toys and other similar devices around my house.