Very happy to share our new work APRL (+ open-source code release)! The important step forward we took here is enabling the robot to keep improving with more data—walking faster and adapting to new situations—where prior work saturates.

Navigating the trade-off between efficiency and performance is tough but important for real-world learning. APRL modulates the robot's exploration based on a notion of 'familiarity', i.e., the robot can explore more aggressively if it can predict dynamics and vice versa

Structuring exploration in this way allows us to not only train remarkably quickly in the real world like our prior "Walk in the Park" system but also reach much higher performance with further training, especially in more complex situations like soft terrain/inclines/outdoors

The code ( has all you need to get started (other than the robot itself): reset policy, simulated and real environments, and RL training. If you have a Go1, give it a try! It should take only a few minutes to set up ◡̈

Thanks to Yunhao Cao and my advisor @svlevine! see Sergey's thread here: + the previous "Walk in the Park" work we built on here:

Awesome work! Also, great to see more open source releases for robotics papers. Keep up the great work! 🙂

Really great work! I am wondering how this method can be further extended to ensure the robot has a natural walking gait, rather than a crawling-like walking gait.

Very cool, congrats @smithlaura1028 !

Wow.