Comment by wyager | Hacker News Reader

Comment by wyager | Hacker News Reader

Parent Full thread

> Some kind of machine learning (NN) would probably work well.

This isn't exactly a constructive criticism, but that is essentially the most useless comment you could possibly make here. "This very general class of algorithms that I've heard about a lot would probably work here.". That is true! It's sort of like saying "Some kind of processor (Intel) would probably work well". It's completely incidental to the problem at hand, and most likely not worth saying.

It's also just a bad suggestion, because you almost certainly don't need to resort to the "high-interest credit card of technical debt" for what is a control and kinematics problem. Even if there's not a nice closed-form solution, there are 100 other approaches that are more appropriate for this sort of thing than just jumping straight to a neural network. I'm not going to go into an explanation of control theory, but "use a neural network" is not the universal answer to every problem in engineering and computer science. /rant

You can set up landing from a jump as a two-point boundary value problem. It's like orbital rendezvous - there are both position and velocity goals, and more degrees of freedom in the goal than actuators. This is called an under-actuated system. So you have to exert changing forces and torques during the period of control. This is rocket science, which is good, because rocket science has been well-studied.

I used to work on this stuff.[1] Raibert's insight was that balance dominates gait. My insight was that slip control dominates balance. On flat surfaces, slip control isn't a big issue, but on hills, it dominates the problem.

Here's a simple example of a two-point boundary value problem. Suppose you want to accelerate an stationary object so that at time T1 it has velocity V1 and position P1. Your only controllable variable is acceleration, but you can change that over time. No constant acceleration will do this. You're going to need a changing acceleration. For this particular problem, there are closed-form solutions. Look up "shooting method" for how to solve this in general.

An appropriate use of machine learning here would be to correct for small errors. You can calculate an idealized solution from the dynamics, and when you use it on a real robot, it will probably be somewhat off. That's when it's appropriate to use machine learning, or at least hill-climbing, to tweak the tuning parameters until it's on target.

Obligatory Freefall: [2]

[1] https://www.youtube.com/watch?v=kc5n0iTw-NU [2] http://freefall.purrsia.com/ff2900/fc02895.htm

I think that neural networks work well in muscle memory applications - repetitive and fast motions. I think in this case, the robot could learn (in a similar manner to other systems [1][2]) some fundamental motion characteristics (primitives) for jumping. Then, combining and augmenting primitives in such a manner so as to vary its jumping and landing capability.

So, to me, it seemed that it was the kind of problem where a neural network would probably work well. I never said it was the correct, optimal, best or only answer. Maybe as the jumping capability and requirements become more complex, a neural network would be better?

It is likely that a specifically designed dynamical control system would work better for a limited series of simple jumping/landing tasks, but it would be interesting to see how well a neural network system would run for much more complex or varied tasks.

[1] http://ieeexplore.ieee.org/document/5686298/

[2] http://ieeexplore.ieee.org/document/4650953/

I think this is actually very constructive criticism and needs to be heard more on HN.

For anyone who doesn't get the quote, "high-interest credit card of technical debt" refers to a paper published by Google in 2014.

PDF link: http://static.googleusercontent.com/media/research.google.co...

Please be kind to everyone :) I think there are better ways to state your criticism that don't read so harshly. And in being kind, I think people are more receptive to criticism and suggestions. It can seem pretty silly from a purely rational perspective, but we're all real human beings, trying to learn and understand things together.

While i agree with both the parts. I like to ponder on "high interest credit cards". I believe, hacking in some form is also tinkering with the solved problems with a new approach. The optimal solution might not change but the approach gets refined thus lowering interest rates for something else.

Reply on news.ycombinator.com