Mattias Rost

Researcher and Coder

Actions on Google for my Tesla

Posted on 2020-03-08

I've been meaning to add the possibility to control the AC in my car from Google Home for a while but never got around to it. A few nights ago a finally set to action.

Google Home is basically Google's take on the smart home. It is an ecosystem of things. The google home smart speaker is integrated with the google home app which is integrated with the google voice assistant, etc. etc.. In the Google Home app, you can setup devices you have around the house that are compatible, such as lights, power outlets, thermostats, etc. Once they are setup in the app, you can control them using google assistant or google home smart speakers. So, if I want to turn to be able to turn on the AC in my car by saying "Ok Google, turn on car AC", I must make sure I can have a device in google home with the name "car AC" and when I issue the command "turn on", it will send a signal to my car to turn the AC on. To do that I started looking at Actions on Google.

(Do I really need to build something for this? No, there are a couple of actions on google integrations already that you can simply add. However, when you do, you basically give away all control of your car to a third party. With this control and physical access to the car, you can essentially walk up to the car, unlock it, and drive away. Now, I'm not too paranoid about this, even though I probably should be. But, in this case, instead of giving someone the chance to drive away with my car, I much rather learn how to build this myself. After all, it's really quite easy.)

Actions on Google allow you to "integrate with google assistant". This currently means that you can extend your mobile apps to work with the voice assistant, to create conversational apps (basically apps that you use through voice/text... yes, chat bots), or connect to smart devices (i.e. google home). For my purposes, I do not want to create a conversational app. If I did, I would have to say things like "Ok google, talk to my car. -Ok, talking to your car. -Hi I'm your car what can i do for you? -Turn on AC. -Ok, AC turning on". Zzz. I much rather just open the app on my phone and press the climate button. So, instead I went for smart device integration.

Actions on Google for smart home devices is really quite simple. You need an Oauth server that can issue a token, and an https endpoint that respond to commands issued by the google assistant. We are simply building a service that acts as a server for a particular device type. There are only four commands possible: Sync, Query, Execute, and Disconnect. WIth an access token received from your Oauth server, the Sync command should respond with a list of devices that this user possesses. Query should respond with the state of the queried devices. Execute should abide to whatever the particular command is (e.g. TurnOn should turn on the device), and Disconnect should do what ever needs to be done whenever the user unlinks their google home with your service.

So I first implemented the Oauth server. There are essentially two routes required: auth and token. The auth route should let a user authorize themselves and tell Google that it is ok for your service to share access to their devices. The server responds to google through a redirect route including an authorization code. The token route should take an authorization code and return an access token (as well as a refresh token that can be used to refresh the access token). For more information go here.

In my case I let a user authenticate with their tesla account, and use their access token as the access token (together with some other stuff). This means I never have to bother storing any information on my server, but let Google store that in the token instead.

Next, the action route. Google issues POST requests to the route with access token in an Authorization header. The post data is JSON and includes a requestId and a list of intents. The intent specifies if it is Sync, Query, Execute or Disconnect.

Sync does not send any additional data, but simply expects to have the devices associated with the user returned. For this, I lookup the devices associated with the tesla account (as given by the access token) and return the device data mapped to data expected by Google, including an id that can be used to identify the device when issuing commands to the Tesla API in the future. For each device Google also expects the type of device, and its traits. The device type is not really important as far as i can tell although it will dictate what it looks like in the Google Home app. But the traits specify what commands can be sent by google assistant. In my case I want to be able to the the heat on or off, so I picked the device type action.devices.types.HEATER (CAR doesn't exist as a type), and the trait is actions.devices.traits.OnOff.

For Query, Google sends a list of devices it wants the status of. Here I simply iterate through the list, and fetch the status of the climate of each device. (The Tesla API has four distinct states fetched through four different routes: climate, drive, vehicle, and charge.) With a device trait of OnOff, Google expects a state named on to be true or false. I set it to true of the climate is on.

Finally, (I don't bother with Disconnect as there is nothing to do), I implement Execute. Execute is a bit of a handful as it will send a list of commands as a set of devices and executions. Here I cheat majorly and assume there will be only one command, with one device, and one execution. The execution of action.devices.commands.OnOff carries a parameter of on that is set to on when the command is to turn the device on. Depending on this value I either turn the climate on or off through the Tesla API. Finally, Google expects you to respond with the final status of the commands for each device.

Finally, when setting up the Action on the Actions on Google Console, you simply add the auth and token routes for your Oauth server, and the action route for your Action. When complete, hitting test will set the action up for you to use through your own google assistant only.

All in all, there are a few concepts required to learn around Actions on Google, but once you understand those, enabling you to control something through google assistant is really as simple as writing a small rest API. Next I might add more capabilities, such as adding both the car heater, and the car charger, so that I can start or stop charging at home. I might also add the capability to set the temperature I want in my car, although I very rarely change this so it's very far down my todo list.

And here's the code.

I am an academic again!

Posted on 2019-07-18

After a little more than two years at Iteam, I am now back in academia! This week I'm starting my position as a senior lecturer at the University of Gothenburg. I will be affiliated with the division of human-computer interaction at the department of applied IT. The division is just starting as I join, so I will be working together with Alexandra Weilenmann to establish it.

I am currently interested in understanding how mobile technology (and IT in general) is affecting our well-being (both positive and negative), and will continue looking in that direction. I'm also very keen on collaborating on this with other researchers, across disciplines so please reach out if you find this topic interesting.

I am excited about being back as an academic, both to do research and to teach and inspire students.

Template for PoC using NodeJS, React, and Express

Posted on 2019-01-05

Quite often I need to try a simple idea for a new service or app. My weapon of choice is currently nodejs and react. It is fairly easy to quickly hash out a server backend. Using create-react-app or NextJS, it is also quite easy to write a react frontend for said backend.

This is all nice and well for a bit while you try out the idea for yourself. But no idea is really worth anything until it is shared with others. So now you need to host both the frontend and backend somehow, and point the frontend at the backend. This typically involves too many steps to make it feasible to try out the very simplest ideas.

What I typically want, is a single server that can run somewhere that hosts both the backend and the frontend. Ideally just a Dockerimage. And I want this to just work without bothering with builds and such. After having done this for a few projects recently I decided to create a template so that I can reuse it whenever I need to try out the next idea.

The result is in this GitHub repo.

To use it, copy the repo. Run npm install and then npm run dev. (I use node 10 and there is a .nvmrc in the repo.)

You now have a server on port 3000 that will restart when you change the server files, and will rebuild the client when you change the frontend. You are basically set to start materialising your idea.

Once you want to share it, you can either use ngrok, or build a Dockerimage using the Dockerfile in the repo and host where you typically host things. When the image starts it will build the frontend. I do it this way instead of when the dockerimage is built so that we can set environment variables on the hosting platform easier.

The repo sets up webpack 4, babel 7, react, express, and jest. The server runs in development mode or production mode. In development mode the server restarts when the backend code changes, and rebuilds the frontend using webpack on client code changes. In production mode the server builds the frontend on startup. It is called production mode, but I would not recommend this for actual production ready applications. But for a PoC this works very well.

Feel free to open issues or send PRs on the github repo.


Low power sensors using nRF24L01+ and ATTiny

Posted on 2018-11-27

Smart home devices are expensive. A simple temperature sensor costs quite a bit. If you want one in each room it quickly gets costly. If you want door sensors on all doors and windows it's pricey.

Cost is not necessarily a problem. If you know what you need you can make a budget. But if you don't know what you need then it stands in the way of experimentation. I want to be able to put up a bunch of stuff in the house and see what I can do with it. It's hard to justify that cost for other than oneself.

So I've been on and off playing around with different ways to create cheap sensors. Not just to have these sensors, but because it is fun to learn.

This is something of a report on my latest adventures within this.


I started out with Arduino pro minis as they are easy to program and test things with. And it all starts with communication. My thought of sensor network consists of nodes that are sensors and/or actuators (think lights and motors), and gateways that can receive sensor readings and send commands to actuators. Since cost has been an issue, I've come to fall in love with nRF24L01+ - super cheap and low power radio modules.

Arduinos are super convenient, but they are entire boards. You can do a bunch of stuff to get them power efficient. However my thought is that all I really need is the MCU. So instead I want my sensors to run on ATTiny84 chips. They are a little harder to flash, but you can use the same Arduino tools as for an Arduino. Libraries even come with attiny compatibility so it's hardly any difference from a software standpoint. You have less memory but these sensors only read a value and send the value to the radio module so I don't need much memory.

Getting radio to work

The nRF chip communicates using SPI. It requires no more than 3.6V. More and they break. It draws very little current when it is idle, like microamps. But when it wakes up to send data it quickly draws around 15mA. It means the power source must allow quick power draw without causing a massive voltage drop. A capacitor helps. (The capacitor acts as a low pass filter, smoothing out a potential voltage drop.) In my setup I power the sensors with either 5V from wall connected DC transformer, or by battery. I therefore use a 3.3V pull-down component, and a 0.1uF capacitor. I'm not sure I actually need this when powering with batteries and will test this out soon. I use two AA as batteries. I've been successful using CR2032 occasionally but it often fails. I think I need better capacitors. A CR2032 does not have a high peak power draw and can't deliver enough power when sending without too much of a voltage drop (causing the radio to stop working). A bigger on like CR2450 should work but haven't paid my hands on any yet.

Powering the device is the main thing. After that comes wiring and software, but they are easy. I use the RF24 library by TMRH20 for software. It details the wiring. Make sure you get the CE and CSN correct. I've flipped those at several occasions.

I've had a lot of issues with getting Ack messages through. I think it has to do with power source and can possibly be fixed with capacitors. Still need to work on this. The data gets through, but it would be nice to know if data gets received or not. Especially for actuators. Most sensors will send it's data again in a while so a drop of message is no biggie, except for door sensors that I'd like to resend every second until received by a gateway.

My setup consists of an Arduino pro mini communicating with a computer over serial. Essentially byte data received by the Arduino is sent in ASCII over the serial (simplifies debugging). Received data is published on mqtt to be subscribed on. The Arduino also reads from the serial and any data received is sent out over radio. It takes address and 32 bytes of data. There is very little logic on the gateway.

The idea is to be able to have several gateways around the house. Using raspberry pis would make them cheap. Using esp8266 or esp32 would make it even cheaper.

The gateway computer code is written in node.js.

Obviously I can connect the nRF chip directly to the gpio on an RPi, but this is simpler. I also have a bunch of computers laying around acting as gateways now.

Getting them to sleep

While power draw is low on the radio when idle, the MCU still draws power unless put to sleep. (Otherwise it just churns away at whatever clockspeed you've configured, looking for something to do. Delays are merely loops doing nothing but counting clockcycles.) So, we must tell it to go to sleep, getting it to use as little power as possible, and still be able to wake it up to occasionally do something useful.

When sleeping the sensors are barely using any current at all. Somewhere I read people saying that similar setups use less current than the battery looses by just laying around. This would mean that e.g. a door sensor on a window that never opens will work will function for the duration of the lifetime of the battery. That is quite impressive.

However, if this is the case, how will I know if the sensor is still working? To fix this I'm planning on also waking it up occasionally and send a heartbeat. That way I will be able to detect when a gateway haven't heard from a sensor in a while and act accordingly.

Waking up

Getting it to sleep is fairly easy. Set some bits in some registers and off to sleep it goes. Getting it to wake up is done through interrupt. An interrupt is something that allows an MCU to halt it's normal control flow to do something different for a while. It also happens to wake up a sleeping MCU.

I use two types of interrupts in my sensors so far. First one is a watch dog timer. It essentially let you specify a duration in which you want to be interrupted. It can be anything from a few milliseconds, to a few seconds (8s on the attiny). The second one is Pin Change interrupt that will alert the MCU that the logical value on a pin has changed. I use these in essentially the same way: I setup the interrupt I want to use, power down the radio, and go to sleep. When it wakes up I power up the radio, and continue the program which typically involves reading the value of an external sensor and send that to the radio module.

An important lesson I learned recently was to implement each ISR callback for each type of interrupt. E.g. there are two pin change interrupt handlers, PCINT0 and PCINT1, and one for WDT.

In the case of temperature sensor I ask the cpu to sleep for 8 seconds. I do that 6 times, and then send the temperature. The reason is that I don't need to know the temperature every 8 seconds. Once a minute is fine. 6*8s is close enough.

On the door sensor I set the interrupt to react to the door pin - the pin that is high when the door is open and low when closed. This means that the sensor is basically sleeping almost all the time.


So far I've got very power efficient nodes for temperature and doors. They send data to a gateway that publish it over mqtt. I've implemented sending data to nodes, but have no nodes setup for it yet. Xmas lights might be a candidate as I have a bunch of WS2812B lights lying around.

I recommend anyone being interested to look at

For the code for all this and more details, checkout my repos:

Nrf-gateway - nodejs server gayeway
Nrf-nodes - Arduino code for the nodes addon for backup to S3

Posted on 2018-11-18

I've been playing around with Home Assistant lately. I was surprised to not find a way to make external backups automatically. Most what I read recommended making snapshots and copy the snapshot over a samba mount manually. I wanted a way to directly make a backup and upload it to AWS S3. So late last night i threw an addon together. It can be found on github here.