Science fiction might be filled with humanoid robots that look and move like us, but it turned out that designing a robot to walk on two legs was an extremely challenging process.
Keeping balance while moving on two legs needs constant tiny adjustments, and this can be very difficult, especially on difficult terrain or when experiencing unexpected conditions such as being pushed.
This isn’t to say that impressive bipedal robots do not exist.
Boston Dynamics has even showcased humanoid robots that can breakdance, but they still need meticulous programming to provide the choreography and do not deal well with the unexpected.
Now, a team from the University of California, Berkeley has allowed a robot to teach itself to walk in pretty much the same way that a human does – by trial and error.
Robots tend to be far more prone to damage from a short fall than a toddler, however.
They are also extremely expensive to develop, so letting the machine repeatedly fall over in real life wasn’t an ideal situation.
Instead, they first trained the robot in not one, but two different simulation programmes.
Robot learned to walk in two training environments
The first was a large training environment and physics engine called MuJoCo, which stands for Multi-Joint dynamics with Contact.
The system was able to try various movements and strategies within this environment, learning what worked and what didn’t through trial and error.
It was then transferred to a second simulator called Matlab SimMechanics, which brought in more real-world elements for the algorithm that would eventually control the robot to learn to deal with different conditions in safety.
Eventually, after extensive reinforcement learning, the algorithm was transferred to the robot, which the team had named Cassie.
The machine itself is very much bipedal, consisting mainly of a pair of legs, and was put through its paces in the lab.
Impressively, it required no real-world fine-tuning as it was immediately able to walk in different directions and at different speeds, squat down, deal with different surfaces underfoot, and even right itself when pushed with a stick.
The team claimed that the results were better than programmed-in alternatives.
More importantly, this approach provides a versatility that could eventually lead to robots teaching themselves to complete physical tests from the safety of a virtual environment.
Today’s news was brought to you by TD SYNNEX – the UK’s number one solutions distributor.
Read more of our latest Industry Updates stories
