Hacker News Reader

Comment by karmicthreat | Hacker News Reader

It has a great kinematic system. But with all those QR codes scattered around it obviously has little knowledge of its environment. I think this understanding and planning is going to be the big hurdle for Atlas style robots rather than just walking around.

Just look at the number of types of door out there and what it would take for a robot to figure out how to navigate it without having encountered the exact door before. Its going to need to figure out that it is a door. Experiment to find the unlatching mechanism. Figure out if the door swings or slides and in what direction. Then open it without hitting its self and finally make its way through. Then exporting the training data so other robots can learn it as well.

View on HN

For anyone interested, the qr code system is AprilTags: https://april.eecs.umich.edu/wiki/index.php/AprilTags Works like QR codes but is optimized for CV tasks and works well on embedded systems.

It seems that this demonstration was for a factory robot, so dealing with all types of doors is not required. Operating in a factory is much easier than being able to adapt to various settings.

I'm a 38 year old human and I occasionally try to push a door that should be pulled. Getting a door wrong the first time isn't holding me back that much. The same goes for a robot. The main difference is the the robot can go through the door without opening it if it really wants to.

I'm sure they're working on getting that integrated with Google Cloud Vision api to recognise objects without the qr codes :)

https://www.youtube.com/watch?v=eve8DkkVdhI

The QR codes probably encode behavioral scripts to apply to an object that the robot encounters. So a box with one QR code may mean that the robot should attempt to pick it up even after interruption. A different QR code may mean that the robot should give up if it is interrupted when picking up the box. A different QR code may mean pick up box and move it to self. A QR code next to a door may mean that the robot should exit it automatically when it gets within a certain distance of it.

very unlikely. they were probably used for localization. for instance, locating door for egress, or computing a grasp for the box. grasp planning is particularly difficult at the moment.

I think it could have dual purposes. There was an earlier segment of the video where the robot was able to find and pick up a box and stack it on a shelf without the presence of the QR codes. There are a lot of clever uses you could put to QR codes as object and environment tagging. It would be effective at encoding what to do and perhaps what can be done with the box. I used a similar strategy to define behavior scripts for the Aldebaran NAO. I printed out some barcodes and taped them to the walls so that if the NAO was exploring an area and saw certain barcodes on the walls, it would do different things based on the codes that were present. This way, I could do simple behavior scripting without having to change the robot's programming.

for industrial settings, i'd imagine they'd (start to) optimize the interior design for the robots capabilities rather than for human aesthetics. Less (or different bandwidth) lighting, doors that communicate with the robot, etc.

Reply on news.ycombinator.com