Lynch et al. 2019 introduced “Play-LMP”, a self supervised method that learns to organize play behaviors in a latent space, then reuse them at test time to achieve specific goals. Combining self-supervised control with a diverse play dataset shifts the focus of skill learning from a narrow and discrete set of tasks to the full continuum of behaviors available in an environment. They found that this combination generalizes well empirically—after self-supervising on unlabeled play, their method substantially outperforms individual expert-trained policies on 18 difficult user-specified visual manipulation tasks in a simulated robotic tabletop environment. These studies asked adults to generate play behaviors. However, children appear to be designed to be playful and exploratory in ways that especially enhance learning (see Gopnik 2020 for a review). Our plan is to put both children and adults into a similar environment that is designed to be more playful and child-like in order to generate a new set of training data, and to see if it will have similar effects in training and evaluating the robot (image 1). We will have both children and adults generate the training data. Our hypothesis is that training the robots on children’s spontaneous play behavior in this environment will lead to more robust and generalizable results than training on the adult data set, and that even adults may produce more varied and informative data in the playful environment than the standard one. We were not able to conduct our in-person experiments in year 2 due to the Covid-19 Pandemic and now have received IRB approval to begin testing in Fall 2021 and are extremely excited to finally do so. We have set up an entire VR testing environment in the Gopnik Lab using the HTC VIVE wireless head mounted system and have received IRB approval to test children 8+.
Eliza Kosoy (PhD student)
Pierre Sermanet (Google Brain)
Corey Lynch (Google Brain)
Igor Mordatch (Google Brain)
Prof. Alison Gopnik (UC Berkeley)