Overview

• Footage released Aug. 28 shows the quadruped executing repeatable backflips, culminating in a seven-flip sequence highlighted in the video.
• Robotics engineer Arun Kumar says the objective is to push the platform to its hardware limits to reveal balance failures and boost robustness.
• The team trained the maneuver using reinforcement learning in simulation that rewards successful actions before transferring the behavior to real robots.
• Engineers deploy sim-trained skills to hardware, log failures, and iterate to refine stability and recovery responses.
• Boston Dynamics says customers are unlikely to need flips, with the research aimed at ensuring reliable operation during tasks like carrying heavy payloads.