it has a simulated camera
to run it, you need pyaogmaneo, opencv, and of course Donkey-Gym
Do you think that we can use it ?
also, it uses a modified box2d. I do not understand why though - the original python version didn't modify box2d as far as I know
Some environments of OpenAI Gym are implemented in C++
most stuff is in Python unfortunately
Does it include training phase?
I can try training it for longer, but the motors get really hot on this robot
super!
@Thanh-Binh new video recall demo (with audio): https://www.youtube.com/watch?v=64xac-fITd0&feature=youtu.be
Hey Eric saw the video and looks interesting. I have a few questions about it. When we say recall what exactly is it recalling and is this something stored in the weights or offline in a file or what not? Was also wondering what the difference between AOgmaNeo vs OgmaNeo2? Is AOgmaNeo a different architecture or just stripped down to run on a micro-controller? Thanks.
@Thanh-Binh yes, it uses a "dual encoder" now - it's really 2 encoders with different strengths (one is generative and the other discriminative). The generative one is the older ESR encoder, and the new discriminative one uses prediction errors to update its representation. Individually they are not good on all tasks, but together they complement each other and provide superior performance.
it can now learn to add numbers with a single layer, and still perform long time horizon predictions
It also seems to help the reinforcement learning as well, better performance all around
by itself, the older ESR encoder pooled different inputs pretty well but wasn't able to discriminate certain inputs. The error-driven encoder could discriminate really well but could not pool inputs together well, resulting in a very noisy representation. Together they augment each other and it gets the best of both
Since it is 2 encoders now, it is slightly slower than before. But I think it is worth it
could you pls update a C++ example?
thanks
@Thanh-Binh Guide is now up! https://github.com/ogmacorp/AOgmaNeo/blob/master/AOgmaNeo_User_Guide.pdf It doesn't include code examples, it just describes how the system works briefly, so that one knows what the different concepts are. I will update the C++ and Python code examples next to show how to use the library (C++ first, since python still has working examples)
@Thanh-Binh C++ demos have been updated to run, but are not well documented yet: https://github.com/ogmacorp/OgmaNeoDemos/tree/aogmaneo
It does both of those pretty easily
Unless you mean runtime speed, in which case yes that is expected, it is slightly slower