search

matwerber1 / aws-alexa-smart-home-demo

End-to-end demo of Alexa Smart Home Skill, including AWS IoT Core and physical ESP32 acting as a thermostat
36 stars 10 forks source link

readme

aws-alexa-smart-home-demo

Overview

This project shows you how to build an Amazon Alexa skill to control a (mock) thermostat using the Alexa Smart Home API and an ESP32 development board (optional).

Whereas many Alexa examples focus only on the AWS Lambda function and dummy JSON responses sent to/from the Alexa service, this project aims to give a more complete end-to-end demo by demonstrating ways that one might use Amazon DynamoDB, AWS IoT Core, Amazon Cognito, and other services.

They say a picture is worth a thousand words:

No Physical Devices Necessary

Demo Video

This video shows the ESP32 in action by asking Alexa to change the device's state and asking it for the current temperature: https://www.youtube.com/watch?v=Cc9Y0D2bzJ8

Disclaimer

This is my first dive into AWS IoT + an Alexa skill. There are no doubt ways to do things better; some of my design choices may not be ideal. For example:

Prerequisites

  1. AWS Account with administrative access
  2. A pre-existing Amazon S3 Bucket to store CloudFormation templates (or, you can create one as you go)
  3. Node10.x and npm (to install Lambda dependencies before uploading to AWS)
  4. Optional - an ESP32 and related components, if you want to build the physical "thermostat" - ESP32 Bill of Materials

Deployment

Part 1 - Alexa Skill and AWS Cloud Backend [REQUIRED]:

  1. Create Alexa Skill and AWS Backend
  2. Sign up for your skill
  3. Test your skill

Part 2 - Thermostat with ESP32 [OPTIONAL]:

  1. ESP32 First-time Setup
  2. Build your ESP32 Thermostat (See ESP32 Bill of Materials)
  3. Test your skill and thermostat

Architecture Overview

Amazon Alexa skills

At it's core, an Amazon Alexa skill is simply an AWS Lambda function that gets invoked by the AWS-managed Alexa service when a user speaks to their Alexa. You create a skill in the Alexa Developer Console and then link it to a Lambda function in your AWS account, and your users register to use the skill via the Alexa mobile app or website. You need an OAuth identity provider (IdP) to keep track of your registered users, and for this demo we will use Amazon Cognito.

Custom Alexa Skills essentially allow you to do anything you want when your Lambda function gets invoked. Custom Alexa skills give you more control but require additional development and planning.

As an alternative to a custom Alexa skill, the Amazon Alexa team has created a number of pre-made Alexa Skill Kits and APIs that give you a framework for rapidly developing skills for popular use cases. If your use case matches a skill kit's capabilities, it may be worth trading the flexibility of a custom skill for the time savings offerred by the skills kit.

While you could certainly build a smart home thermostat skill from scratch, we will use Alexa's Smart Home Skill Kit & API in this project to speed up development.

The Smart Home Skill Kit & API includes controller interfaces for common smart home devices, such as thermostats, locks, cameras, lights, and more. For this project, we will use the Thermostat Controller Interface and Temperature Sensor Interface.

AWS IoT Core

AWS offers several IoT services, including but not limited to AWS IoT Core, IoT Analytics, IoT Events, IoT SiteWise, IoT Things Graph, and Amazon FreeRTOS.

In this project, we only need to use AWS IoT Core to create a thing, which is a logical representation of a physical device. Even if you do not build the optional ESP32 thermostat in this project, this project allows you to interact with your IoT thing as if it were a physical device.

A core component of AWS IoT Core is the device shadow, which is a JSON document stored in the cloud and which contains two key components:

When the IoT Core shadow service detects a difference between the reported and desired state in the shadow, it will send a delta message over MQTT to the physical device. In normal operating conditions, we expect that the device will receive this delta, respond by changing its state (e.g. switching on/off, adjusting volume, etc.) and then report back its new reported state to the cloud. Once IoT Core see's that desired state matches reported state, IoT Core stops sending delta state messages.

In this project, we will use the device shadow's reported state to track things such as the current temperature, thermostat mode (heat, cool, or off), and current target temperature. When a user talks to Alexa to change our thermostat settings, Alexa will invoke a Lambda which will then change our device's desired state in the shadow. The shadow will detect the change and send the delta to the physical device to respond accordingly.

Amazon Cognito and Amazon DynamoDB

We will use Amazon Cognito User Pools to manage the users that sign up for our skill.

When a user invokes their Alexa skill, the Alexa service needs to know which AWS IoT thing belong to that user so that Alexa can read from (or update) the appropriate device shadow.

We will store the mapping between Cognito users and IoT things in Amazon DynamoDB, a fully-managed NoSQL key-value database.

Questionable Design Decisions

This section contains design choices that I am unsure of or where I want to further explain my reasoning.

DynamoDB to store user-to-device mappings

I chose DynamoDB because I wanted the possibility of a "many to one" mapping of users to a single device. For example, maybe a household (family or roommates) all want to control the same thermostat using their separately-registered Alexa devices and accounts. If we only needed a "one-to-one" mapping, then it would be simpler to use an attribute in our IoT thing's device registry.

Storing Alexa discovery config in Lambda code

When a user asks Alexa to discover available devices, Alexa will query your backend application (via a Lambda) to ask for a list of devices associated to the user's Cognito ID. You must return a Discovery response in the form of a list for each device containing basic information (e.g. device name, manufacturer name) as well as complex nested structures that describe the device's capabilities (e.g. ThermostatController).

While the simple attributes could be stored as attributes in the AWS IoT Registry, the registry didn't seem like a good fit for the complex structures that describe device capabilities/interfaces. So, I opted to contain both the simple and complex information in a Javascript map object as part of the Alexa Lambda's source code (./lambda/functions/alexa-skill/discoveryConfig.js).

The config object has the following format:

const discoveryConfig = {
    'smartThing-v1': {
        '1.00': {
            manufacturerName: 'SmartHome Products, Inc.',
            modelName: 'Model 001',
            friendlyName: 'Smart Device',
            description: 'My SmartHome Product!',
            displayCategories: [
                'OTHER', 
                'THERMOSTAT',
                'TEMPERATURE_SENSOR'
            ],
            capabilities: [
                {
                    type: "AlexaInterface",
                    interface: "Alexa.EndpointHealth",
                    "version":"3",
                    properties: {
                        supported: [
                            {
                                name: "connectivity"
                            }
                        ],
                        retrievable: true
                    }
                },
                ...
            ]
        }
    }

The first key, smartThing-v1 is the physical device's model number, while the second key, 1.00, is the device's version. The keys within that represent the device's basic attributes and Smart Home capabilities. My thought was that both the physical device (model number) and its firmware version both dictate what capabilities it has, so that's how I opted to organize configuration.

As I revisit this logic, I do like the idea of keeping the complex nested structures for things like capabilities as part of the Alexa skill's source code, as these (should) be static for a given model and firmware version and generally only need to be retrieved by the Alexa Lambda.

However, I imagine that in a production scenario, you could have the same model number and firmware version built by multiple manufacturers, so it feels wrong to include this as a static attribute. I think it would be better to move this to either an attribute on a per-thing basis in the IoT device registry or to a separate data store like DynamoDB.

Modifying Shadow's Desired State from the Physical Device

Most of the literature I read calls for two de-facto rules:

  1. Physical device should only modify reported state and never modify desired state in the device shadow.
  2. Cloud should only modify desired state and never modify reported state; only the device knows what reported state is.

I agree with #2, but ran into challenges with this demo that could only be solved by breaking rule #1.

First, for this project, my functional requirements are such that:

  1. The thermostat should be controllable by both cloud-side and physical button on device
  2. The user's physical interaction with the thermostat should always overrule desired state (if any) previously set by the cloud

The challenge I ran in to was this:

  1. Hypothetically, say a user previously asked Alexa to set mode to COOL; Alexa (via Lambda) set desired state to COOL, the thermostat received a shadow delta via MQTT, and accordingly changed the mode to COOL and sent a reported state value of COOL back to the shadow. Desired and reported state match, we are at peace.
  2. User walks up to the thermostat and presses a button to change the mode to HEAT. The device correctly changes the physical mode and then sends a shadow update of reported = HEAT.
  3. AWS IoT Shadow Service now sees a delta in state, because reported = HEAT but desired = COOL (from previous Step 1)
  4. Shadow Service sends a shadow delta message to the thermostat (desired = COOL)
  5. Thermostat sees that desired state is COOL, changes physical state to COOL, and sends a shadow update of reported = COOL.
  6. Again, reported and desired state are in harmony and equal COOL, but our user wants the mode to be HEAT.

The only way I could solve for the problem above was to have the physical ESP32 thermostat clear the desired state when the user physically interacts to change the mode (i.e. pushes the mode button).

I think the theory behind my solution is correct, but the actual implementation is faulty as I didn't account for scenarios where, for example, the device is disconnected from the internet and the user presses a button. In such a case, the command to "clear" the desired state would be dropped and when the device re-connects, we would again run into the same problem. Perhaps incorporating timestamps into state requests make sense... e.g. if a delta is old (because of previously-lost connectivity), ignore it and clear desired state? Or, if a device was disconnected and the user physically changed its state, compare the timestamp of the physical state change with the timestamp of the desired state change and let the latest timestamp win?

I suppose it depends on your specific use case and who/when/where/how disputes should be resolved. For ecample, if your business use case was such that "orders from the cloud should always override user's physical interaction with the device', then this would be a non-issue. For normal retail and consumer smart home applications, I imagine user interaction should always take precedent. For commercial, industrial, and other large-scale use cases, perhaps cloud should always take precedent?

Would love to hear thoughts from anyone that has production experience with this type of scenario.

User Stories

I've created the Appendix - User Stories to document my (basic) understanding of the stories that a Smart Home Company, device manufacturer, and end-user might follow.

Cost

On the AWS side of things, everything we do should cost near-zero (maybe pennies a month). I haven't calculated exact costs, but the AWS Free Tier covers a lot. As of this writing (July 2019), the "always free tier" provides more than enough usage to cover:

The free tier also includes 12 months of free usage for certain services, including but not limited to:

Even if outside of the free tier usage, everything we are using will be pay-as-you-go based on actual usage, which should amount to mere pennies a month. That being said, remember you are responsible for watching your cost and usage.