9 months ago •
Whenever I have a small number of Reactrons that are meant to work together as essentially a single unit, I call the arrangement a subcomplex. It is a complex, in that it consists of more than one internal Reactron. But it is "sub" in the sense that these internal units exist below a unified Reactron interface for the complex, so from the outside, it may appear as a single unit.
As an example, if you consider the Reactron coffemaker, water pump, and coffee-contextualized simple button, those are actually three separate Reactrons. But, because these Reactrons are a special case scenario, where the button has a one-to-one relationship with the water pump, and the pump has a one-to-one relationship with the coffeemaker reservoir, they could have been structured as a single, three-unit subcomplex, with a single network interface, instead of three. I didn't do it this way because these units were all conceived and added at completely different points in time. Additionally, I like keeping the button and pump abstracted, because if I ever need a general purpose button elsewhere, or a pump, these designs can be replicated easily without de-coupling a coffeemaker. That is one of the basic tenets of Reactron Overdrive - keep units simple and abstracted, and create desired complexity with increased numbers.
In developing the speech-interacting Integron, I have done a lot of testing and analysis, and suspect that the base unit is perhaps too complex. It may be truer to the principles to abstract the speech processing from the human interaction, where the Linux module is a sub-Reactron module in its own right, and the sight, gesture, and audio components comprise a separate sub-Reactron.
I had been working on an "Integron relay" where a subset of the human interaction components were present, but offloaded the heavy processing to a separate unit. My thought was to create a device that could act as a (door) threshold device, like a doorbell intercom, but fully integrated with the whole network. If a doorbell, this would be externally mounted, and therefore subject to weather, damage, potential theft, etc. So it would be beneficial to have it be cheap and replaceable, with nothing critical in it at all. It would just be a dumb terminal, effectively.
After a lot of development on the Integron unit, I am thinking this should actually become the standard model. All Integrons should be a sub-complex, with an Integron relay of whatever form (doorbell, automobile, desktop, wall-mounted, ceiling mounted, wrist-mounted) presented to the human, and all the data processing for speech and most Reactron network functions located on a separate unit. The two sub-units could be physically located together, but need not be. This will allow the use of much higher performance computers to be utilized for speech processing, and will also create better machine utilization efficiency, since we generally do not need as many speech engines as we do interface points. The separation will allow a few-to-many relationship of engines to human interfaces.
It also allows a completely different handling of the audio, escaping ALSA and allowing Linux to just handle received waveform data without trying to play it or capture it. The real question is, will the transport of data to and from the relay units be faster than the capture, processing, and playout all on a local Linux SBC? I don't know yet, but I am going to test it.
In the case of the Automobile Integron, I think the hardware will be pretty much the same, but wired and coded differently. I will still use a BBB for the speech processor, but now the microcontroller will handle the audio capture and playout instead of ALSA. That subcomplex will be entirely local to the car, as GPRS would not be efficient enough data transport, performance-wise, and further, loss of signal would end up disabling...
Read more »
9 months ago •
A lot of the Integron unit is working in component parts. It remains for me to coordinate the various internal subsystems. Here is an image of the first cut of the desktop morphology.
In this image the screen is present but not active, and the little ripple in its surface is just the protective plastic which I have not yet removed. A NeoPixel is activated blue, under an internal diffuser, in front of a sound baffle and support structure joining base disk to top disk. The fabric sleeve is an acoustically transparent speaker grille fabric. There is some left-right asymmetry due to uneven tension, due to me assembling and disassembling this prototype a number of times.
Still working on the mechanical fit and angle of the PIR sensors in the support baffle structure, and I have yet to test the transparency of the acoustic fabric to the ultrasonic sensor. The infrared passed just fine, so worst case if the ultrasonic does not work I will put a central PIR in a tube to narrow the angle, and use that to indicate presence of someone standing right in front of the unit. But I'd rather use ultrasonic ranging since then I get an actual distance.
The unit looks great in a darker setting.
It also is effective at giving a colored light signal from a distance in normal light, though the camera does not report well what the human eye sees, which is much more even than this image shows.
The unit in this image does not contain the Beaglebone Black and audio components, I have that assembled on a breadboard separately for testing, but expect to be assembling it all together shortly. (It does fit - using a proto cape on the BBB.)
The speech synthesis and speech recognition are working. The USB sound card is working well, but I am still working on the right microphone and audio amplifier on the output. Currently I have a capsule electret mic sourced from Adafruit, and it works quite well close up. But I have ordered higher sensitivity ones, because I want to be able to speak to it from a distance, and it's just not there yet. I have also ordered some adjustable pre-amplified mics, which may be OK if I attenuate the signal all the way down so that the peak-to-peak does not overdrive (!) the mic input. The sound card is apparently able to withstand abuse - I did try a sustained, full 2 volts peak-to-peak signal to see if it would fry (the sound cards are only a few dollars each, that test was worth it to find out the capability.) Terrible sound, of course, super distorted - but it didn't fry the board. We'll see if we have to go there at all. The internal baffle of the Integron unit is designed to be a sound cone. On the output side of things, I am using an Adafruit audio amp, but while it is quite excellent at what it does, it may not be the best choice for this application. I have some less expensive PAM chips on order, and some different ones I have used before in house, I will be experimenting with them soon. I also changed the speaker to one that was less tinny. The speech output is mono, so I am just using one speaker of a pair. This unit is really a point source so in some ways it makes sense to combine channels for any potential stereo signal sound files and just use mono. That makes it more compact.
The Moteino is driving the NeoPixel without issue, but I will use more than one in the final unit so that the top, middle, and bottom can be illuminated differently to give three different levels of signal. (That is the idea anyway, we will see if it is practical.) Due to the internal baffling, there may need to be three LEDs per tier to enhance visibility from all sides, so nine total. I am considering making the screen bezel ring out of translucent white acrylic, and adding another NeoPixel to make that a soft, power-on indicator light (overridable of course, if you want it off). So ten RGB LEDs. In...
Read more »
9 months ago •
I've been holding off on an official component list until I can stabilize the build a bit more, but I wanted to mention a few things that have occurred since the original plan was hatched, and some of the design decisions involved.
Without delving too much into the history, I will just say that at one point I was coding a very simple speech recognizer to run on the ATMega328P, and I was able to create some code that was about 85% effective at recognizing a handful of carefully selected keywords that were pretty distinct in their pronunciation. That was the true positive rate. There was also a high-enough-to-be-annoying false positive rate, which would trigger my “initializer” (more on that at a later time). Anyway, I was getting sucked into the details of speech recognition and optimizing for a tiny processor, and this is really not the area where I can add the most value. So, I moved to R.Pi, to try to implement open-source speech recognition using Pocketsphinx, like many other builds. It meant having to use an SBC w/ Linux in addition to the ATMega328P based board, but for this usage I was not so space-constrained so that was OK, even with the added cost and power usage. Also, moving to Linux enabled speech synthesis as well, using festival. Before, I was playing out pre-defined sound files containing sounds and canned speech responses. But I wanted a more general interface, one that could recognize more than a handful of pre-defined keywords, and trigger more than a handful of pre-defined sounds. Ultimately I moved on to BeagleBone Black as the SBC for performance and I/O, but this is just for these Integron units. Other Reactrons that don’t require all this library support are fine on R.Pi, or just stand-alone ATMega328P, or even stand-alone Android devices, based on the specific application.
A Reactron is defined by its communications protocol and internal data structure, not by its specific hardware complement. I wanted to mention that while my header image shows a panel of ATMega328P boards (based on the Moteino by Low Power Lab), Reactrons are not Arduino clones, though their hardware complement may contain one, several, or none. I chose the image because to me, it evoked “many small computers working together”, which is what the Reactron network is all about. In my other images, you will see a multiplicity of BBBs and R.Pis and other hardware, as the mix of hardware for a generalized Reactron completely depends on its purpose. For instance, some existing Collector units are purely BBB boards with attached sensors, and the software to support the Reactron interface. Some Recognizers are just BBBs running statistical calculations on inputs from Collectors - without any directly attached auxiliary hardware. I haven't had a chance yet to write much about Recognizers, but that is coming soon.
Moving on to some of the changes, specific to the Integron unit:
One excellent thing that has changed since I started this project is that the Moteino has been upgraded. [Felix] at Low Power Lab has informed me that all new R4 Moteinos are now shipping with the MCP1703 voltage regulator. This is good news for me, because now I am able to just add “Moteino” to the components list, instead of a board "based on the Moteino", where I would have to do have a separate BOM and so forth. Now, you can just use a stock R4 Moteino (with HopeRF radio and a 4Mbit flash chip) for any of my projects that require an RF enabled ATMega328P-based solution. At the time I started this project, the Moteino was shipping with the MCP1702 voltage regulator, and that meant that using a then-stock Moteino in my 12+V projects (like the Reactron water pump or the newer Automobile Integron) might have resulted in a puff of smoke and little else.
I would also like to direct your attention to the fact that [Felix] has entered THP with the Moteino Framework, and you should go there and give him a skull. [Felix] is simply amazing and creates an enabling technology...
Read more »