Introducing a dynamic sensor network architecture driven by AI. Instead of static sensors, SMOVE will deploy roaming autonomous sensor platforms which collect sound, visual and connectivity information.
Our current prototype uses a scooter as a vehicle to teach our AI software how to navigate paths via human riders. With self-redistributing, orientating, parking and charging features, SMOVE will also provide an e-scooter sharing service seeking to solve the last-mile problem faced by the average commuter conveniently and economically, enabling a car-lite society.
A dynamic sensor network holds several advantages over conventional static sensor networks. Instead of a dense array of low-cost sensors, we can integrate a few high-quality sensors in a mobile platform to cover a large target area. Instead of sending maintenance teams to each sensor location, we can perform preventive maintenance centrally. These ensure competitiveness in terms of cost per unit area. We can also collect data exactly when and where it is required. For example, after a concert, we can send in scooters to collect litter information and provide transport for concert-goers.
In particular, SMOVE approaches data collection innovatively. SMOVE provides a monetization opportunity as sensors can be integrated with an autonomous e-scooter sharing service. Leveraging higher-power sensors like cameras and powerful processors already in these e-scooters, SMOVE incorporates sensor fusion and AI to infer the context to meaningfully improve urban living.
These are made possible because our scooters autonomously visit charging stations to top up onboard lithium batteries, resulting in battery life exceeding a year.
We have chosen the Snapdragon 845 SOC and the Hexagon 685 DSP which accelerates neural net inference as compute. The BM160 Gyroscope, AK0991x IMU and GPS to enable precise localisation, while integrated X20 LTE modem and 2x2 802.11ad antenna configuration allows real time data upload.
For low-level controllers and custom peripheral connections, we chose the Arduino Mega. It has a real time clock to operate stepper motors and respond quickly to commands from the smartphone. Also, it interfaces with the VESC, an open-source motor controller, to send commands and receive telemetry data.
We have chosen TensorFlow Lite as it has the most support for deployment on mobile devices. In addition, low latency is vital, thus the compute must be done locally. However, running separate modules for obstacle detection and route planning would be too computationally taxing for a phone to handle. Thus, we chose an end-to-end neural network so training can be done beforehand.
We have decided to utilize a 5-layer CNN with the inspiration from an Nvidia as it has been successful in mimicking human behavior. We believe machine learning is the way to full autonomy as it allows every edge case to be considered.
Choosing smartphones as our compute enables scalability. In 2018, 1.42 billion smartphones were shipped, providing an efficient supply chain to tap into. For our custom parts, we are talking to an e-scooter manufacturer in China to mass produce our scooters. We have also contacted PCB manufacturers for production of our custom PCBs.
SMOVE aims to be financially sustainable. The projected production cost of one unit of SMOVE is USD773, excluding an estimated recurring monthly cost of USD81. In our survey with 60 respondents, 93% are willing to use SMOVE, with 79% willing to pay USD1-1.50 for a ride. Based on internal calculations, the projected monetization of USD220-360/month/scooter will ensure long-term sustainability of data collection.
Although autonomy is difficult, we have a 3-step plan. Firstly, “platoon mode” will be implemented where multiple scooters will physically trail behind users to redistribution points. This is easier to achieve as no AI is required for direct shadowing. Next, we trial the solution in a controlled environment - university campuses. After the scooter can autonomously navigate the campus safely, SMOVE will roll out in phases over the next 2 years. Finally, AI will continue running as users scoot. Significant deviation between users’ path and projected path triggers data collection for further training, allowing us to cover edge cases and achieve full autonomy.
SMOVE benefits the community and individuals in three aspects: improving connectivity, helping to reduce visual pollution and improving the sound quality in the community. SMOVE’s onboard modem allows us to track cellular coverage and WiFi interference. Heat maps of cellular coverage equips internet service providers to modify their broadcasting strength. In addition, WiFi channel interference information enables businesses and homeowners to optimise their WiFi network settings to minimize interference and maximize throughput.
SMOVE tracks visual pollution through machine learning on camera data. Real-time inference and classification can be used to identify objects of interest like street litter and graffiti. One application is to allow data-driven cleaning of public areas, with our scooter identifying dirty areas for prompt follow-up by government agencies.
SMOVE also records sound using our smartphone’s microphones. Using machine learning, SMOVE can classify incoming sounds into various categories like traffic, construction and more. Individuals and companies can use mapped noise information to make informed real-estate decisions.