Comment by deutschepost | Hacker News Reader

Comment by deutschepost | Hacker News Reader

Parent Full thread

During university I was part of a robocup team and this is the first time I was genuinely impressed with movement of these robots. These kind of robot is often seen in the Humanoid League and during games they don't move nearly as fast and controlled. If even at all.

That said, during competitions it is not allowed to use external motion tracking as it is used in the paper. The necessary image processing takes a good chunk out of the processing power budget and the quality of image sensors this size doesn't help. Also most teams wouldn't train their expensive robots this way because the risk of damaging it is way too high.

I think this is a glimpse of what's to come in robot soccer. But we are a long way away for self contained autonomous robots to behave this way.

> Also most teams wouldn't train their expensive robots this way because the risk of damaging it is way too high.

They train it in a simulation, and these robots are quite expensive too. The only problem is getting miniaturized hardware, which we are bound to see very soon as there are vast practical applications, I'm thinking dishwashers and refrigerators.

training it will be way harder and the net probably needs to grow to eliminate external motion tracking, but I don't think the image processing and quality of image sensors is the bottleneck. These robots need the right hardware for quick inference, maybe some specialised inference chips or something, but I think if you manage to train a net that can do this, you can deploy it in principle on the robots. It looks like it has a "Intel Core i3 processor dual core" which is definitely not the right hardware for efficient (throughput + power consumption) inference.

The problem is imho that the problem is way harder and we will probably struggle to train a net that can do this in simulation. It might just be impossible with our current capabilities, we might have to wait a few years until research catches up with the problem. But I am not a RL guy and can comment only from my computer vision-perspective.

Reply on news.ycombinator.com