We collected almost 16,000 (5,000 shown) balls from human play, ball throwers, and the robot itself in the real world, and augmented them with random noise. The data is split into serves and rallies. Check it out on @github.com: github.com/google-deepm.... Happy to serve up some training data 🤖🏓.

This dataset of ball states contains the position, velocity, and angular velocity right after a hit. We hope that by sharing these, along with a MuJoCo simulation that uses fluid dynamics to reproduce the balls, we can encourage further research. A big thanks to the @mujoco.bsky.social team!