Scientists Teach Jumping Robots to Learn from Failures and Land Safely
When it comes to robotics, we have seen it all, particularly when it comes to jumping.
We have been privy to Salto that can precision jump like a gymnast. We have witnessed Boston Dynamic's Atlas robot jump and backflip. We have even come across tiny microbots that jump around in your colon to deliver drugs.
All of these bots had one thing in common, however: they do not always land on their feet. Now, a new video has surfaced of scientists teaching robots to jump as high as possible and land firmly on their feet.
They achieve this noble task by using Bayesian optimization with unknown constraints. This specific sequential design strategy for global optimization of black-box functions does not assume any functional forms and is most commonly used to optimize expensive-to-evaluate functions.
In the scientists' preprint, the experts explain how in the past decade, numerous machine learning algorithms have been shown to learn optimal policies; however, failing scenarios have been a common pitfall. Because in robot applications, many algorithms struggle with leveraging data from failures, and this team has something to offer regarding this issue.
The researchers designated failing behaviors as all "those that violate a constraint and address the problem of "learning with crash constraints", where no data is obtained upon constraint violation." They then addressed the no-data cases through a GP model (GPCR) for the constraint that combines discrete events (failure/success) with continuous observations. This enabled them to demonstrate the effectiveness of their framework on simulated benchmarks and on real jumping quadrupeds.
The scientists then reported a positive result outperforming both manual tuning and GPCR. The video of these results (included at the bottom) is fun to watch as the robots wobble but eventually land on their small feet.