On the 28th, Boston Dynamics announced that its quadruped robot, Spot, has achieved complex maneuvers such as continuous backflips. The core technology stems from millions of reinforcement learning training cycles. The R&D team revealed for the first time that initial simulation results, when transferred to the real robot, "failed almost every time." Engineer Arun Kumar admitted, "We fell hundreds of times on gymnastics mats before we dared to let it try on concrete."
The seemingly showy backflips are actually aimed at survivability in extreme scenarios. Official data shows that reinforcement learning has increased Spot's posture adjustment speed by 40% in the event of falls or slips, effectively protecting the tens of thousands of dollars worth of sensors on its back, such as the LiDAR. An unexpected bonus is that the robot's gait has become more similar to that of a real dog due to optimized dynamic control. Joint swing has been reduced by 15%, limb coordination on uneven terrain has been improved by 22%, and it can even maintain balance with rollers installed on its front legs.
This breakthrough continues the Spot family's enduring legacy: from remaining steady after being kicked in 2015 to now performing a backflip to save itself, Boston Dynamics is transforming "trial and error training" into a commercial advantage. It's been revealed that the new algorithm has been applied to high-risk scenarios like oil pipeline inspections. One energy client reported a 67% reduction in Spot's accident rate in slippery pipelines. "Backflips aren't the end game," Kumar said. "We're teaching it to 'predict falls,' so it can protect itself better than humans."