Why So Many Robots Struggled With the DARPA Challenge

stowie writes: The DARPA Robots Challenge concluded recently, and three teams were given prizes for completing all the tasks. The other robots in the competition struggled — not only were they unable to complete the required tasks, many of them were unable to even stay standing the entire time. So why did these robots have such a hard time? "DARPA deliberately degraded communications (low bandwidth, high latency, intermittent connection) during the challenge to truly see how a human-robot team could collaborate in a Fukushima-type disaster. And there was no standard set for how a human-robot interface would work. So, some worked better than others. The winning DRC-Hubo robot used custom software designed by Team KAIST that was engineered to perform in an environment with low bandwidth. It also used the Xenomai real-time operating system for Linux and a customized motion control framework. The second-place finisher, Team IHMC, used a sliding scale of autonomy that allowed a human operator to take control when the robot seemed stumped or if the robot knew it would run into problems." If nothing else, the competition's true legacy may lie in educating the public on the realistic capabilities of high-tech robots.

Comments

Deep learning

[Lennie]

Yes, I'm sure the DARPA challenge is hard work, but I was much more impressed by how well they were able to apply deep learning for use with robots:
http://newscenter.berkeley.edu...

The fastest robot on the DARPA challenge took 45 minutes, look at how fast the robot is in the above video. It's much more close to how a human would do it.

5 years ago from the same lab they took hours to do things and they were still using very little machine learning in comparison:
https://www.youtube.com/watch?...

And more importantly how close they are to using demonstrations (how about YouTube videos or from other people or robots doing similar tasks) to get robots to learn faster and many more tasks:
https://www.youtube.com/watch?...

I was also very much impressed the first time I saw what Deepmind had done:
https://www.youtube.com/watch?...

Re:They all struggled, at first.

[Anonymous Coward]

Google's self-driving cars have logged plenty of time in traffic on public roads, so I don't know where you get the idea that nothing has happened outside of controlled conditions.