Members | Sign In
Legacy MOVI User Community Forum (readonly) > MOVI Question & Answers
avatar

Training Sentences - Automatic training or Low-Level Interface?

posted May 14, 2016 19:21:25 by Pritcharduino
Good day!

I have tried to train MOVI with a rather large amount of sentences. By using the low-level interface, I managed to train around 285 sentences. However, I noticed that there was a large amount of memory left to spare, and had decided to train more. After running through the process of resetting my Arduino DUE, resetting MOVI, and re-opening the serial monitor, I typed the 'TRAINSENTENCES' command, and upon retraining, there were only 8 sentences in the vocabulary.

I am wondering why MOVI did not train all of the sentences it had before, and why it erased pretty much all of them. There is still over half the memory left on the MOVI board, and the sketch only uses 20% of the memory on the DUE.

Another problem I've been having, is that MOVI will train very odd amounts of sentences, and sometimes, not train any at all. I kept waiting for the auto-Train to kick in, but nothing ever happens. When I use the Low-level interface, it only trains around 8-16 sentences. Any idea what the issue is, or how this can be fixed?

Could possibly be that MOVI is not given enough time to train all the sentences. When the board says, "MOVI is loading new sentences", it only trains for around 5-7 seconds, and then stops. What makes no sense, is that in this short time, it managed to train 250+ sentences on one occurrence, and only 8 on another.

Is it possible to provide MOVI with around 30 seconds to a minute of training time? Could this fix the issue with hundreds of sentences?

Thanks!
Building human-scale anthropomorphic robots.
page   1
2 replies
avatar
Pritcharduino said May 14, 2016 19:34:47
EDIT: After leaving my Arduino DUE and MOVI sitting idle for around 30 minutes, I typed the 'TRAINSENTENCES' command in the Low-level interface. I managed to train 726 sentences! There was no lag time or delays between recognition of any of these sentences, so there was no memory overflow on the RAM. I may have discovered that by letting your Microcontroller sit idle, it allows it to "run-through" all trainable sentences, so it will be able to train them for later use through the Low-level interface.

To anyone who reads this: If you use a DUE, and let it sit for upwards of an hour or so, you may be able to train upwards of 1000 sentences, possibly more. I will be sure to keep trying, and see how far I can push MOVI's potential. This looks very promising for a context-driven conversation engine on a robot!

Great work guys on creating this awesome piece of hardware! :)

- Ian
Building human-scale anthropomorphic robots.
avatar
GeraldFriedland said May 16, 2016 03:28:08
Ian,

That's great to hear. I strongly recommend updating your library to version 1.02. I just released a couple minutes ago
(see our blog post). It will probably speed up the process!

Gerald
[Last edited May 16, 2016 16:20:46]
Login below to reply: