Backed by the deep pockets of Google, Waymo recently rolled out its autonomous taxi service in Phoenix. Almost halfway around the globe, and without financial support from a billion-dollar corporation, the Academy of Robotics, an upstart in London, is now testing its autonomous delivery vehicle, the Kar-go. The robotic car will transport medications, contact free, from pharmacies to care homes.
I spoke with serial entrepreneur William Sachiti, Academy of Robotics founder and CEO, who told me about his background in robotics and gave me a virtual tour of some of the Kar-go’s autonomous vehicle (AV) technology.
Sachiti’s AV journey began when he envisioned an autonomous airborne courier—a drone that would provide last-mile delivery services in urban areas. Knowing how much power would be necessary for such a feat, Sachiti devised an innovative network of inductive charging stations placed atop streetlights, allowing the drones to use smaller (and lighter) batteries, leaving more space available for the cargo. An artificial intelligence (AI) system would coordinate where the drones travel, using the tops of lampposts as both guide points and charging nodes. He eventually scrubbed the idea due to residential ordinances limiting vehicle noise to thresholds well below what a drone would produce.
Sachiti realized that autonomous ground-based delivery was the way to go, but he also recognized that the technology was still about five years away.
“That’s when I decided to up my geek credentials,” he recalled, “so I went to university to study artificial intelligence and robotics.”
He attended Aberystwyth University in Wales, whose AI department had expertise in machine vision. (A team of Aberystwyth researchers contributed to the design and calibration of the ExoMars rover’s vision system.) While there, Sachiti grew frustrated with the difficulty of finding a particular book in a massive library, so he and classmate Ariel Ladegaard designed the world’s first AI robotic library assistant. A patron can request a book and then follow the bot as it literally guides them to the book’s location in the stacks.
After leaving the university and continuing his education at NVIDIA’s Deep Learning Institute, Sachiti assembled a team of scientists and engineers and founded the Academy of Robotics. One of the scientists with a background in bio-inspired robotics and computational intelligence pointed out that chimpanzees have a superior ability to memorize visual patterns—one that far exceeds that of humans, despite the chimps having smaller and less complex brains. The team concluded that clever software can outperform complicated hardware, and dedicated its resources to designing algorithms that could run on an inexpensive computer. Another team member had previously worked on vision systems for guided missiles; he headed up the design of a custom microcontroller, which would be built by NVIDIA. The chip is a digital signal processor that’s optimized to run the company’s proprietary operating system and training method that simulates evolution in order to create better AI algorithms.
Meet the Kar-go
The futuristic pod on wheels represents a departure from other AVs, especially on a visual level. But in order to appease UK regulatory agencies, the chassis, suspension, motors, regenerative braking system—pretty much everything but the AV-specific sensors and AI—are standard components used in other street-legal cars.
The AV Tech
The Kar-go uses camera vision to help perceive its surroundings. Ultrasonic distance sensors and infrared cameras provide even more information for the AI to use. The team designed its own sensor to detect the presence of small children and animals. Because the patent is still pending on that sensor, Sachiti was reluctant to share details of how it works, but he did tell me that it uses a combination of cameras and lasers to identify an object and its distance. Frame by frame, the AI uses micro-differences in shading, color and distance to determine an object’s likely behavior based on previous experience. For example, rustling leaves will exhibit a certain pattern of motion, while feet walking will be quite different. The Kar-go emphasizes vision and memory in its pattern recognition algorithms.
The computer and AI software, initially “trained” in a data center before being deployed to the vehicle, do the bulk of the computing on board, using just a few megabytes of memory to hold the operating system. The custom chip includes a microcontroller with a built-in field-programmable gate array (FPGA)—a dedicated circuit that processes data quickly through hardware rather than software. Sachiti described it as “the very definition of tiny AI.” Video information is uploaded to a cloud-based server that conducts an in-depth data analysis and sends the results back to the Kar-go, increasing the vehicle’s knowledge base and enhancing its decision-making skills. Because delivery vehicles often travel the same routes over and over, the Kar-go relies on sensors, landmarks and memory for navigation. Its GPS unit is primarily used for logging where the vehicle has been rather than helping it to find its way.
Training the Kar-go. (Video courtesy of the Academy of Robotics.)
Lacking the financial assets of Silicon Valley, the Academy of Robotics team was forced to innovate. Using limited sensors and a lean computing unit, engineers devised unique algorithms, even earning a patent for their vision system, Bio-sim, which enables the car to adapt to new situations with relatively few training examples and minimal hardware requirements. The company claims that Bio-sim outperforms other more expensive self-driving systems.
A Safe on Wheels
Sachiti likes to describe the Kar-go as a “safe on wheels,” with space for up to 24 packages. Company spokesperson Ryan Maloney elaborated on how the vehicle ensures that only the intended recipient can receive the package: “The hatch works in conjunction with a mobile app. The recipient is sent a token or code which will trigger the release of this specific compartment when they hold their phone up to the hatch. The package management system is designed to reshuffle the packages in transit so that if the delivery route changes and the first recipient then becomes the third recipient, it can still produce the correct package."
The Command Hub
While a human safety driver will be in the Kar-go during its testing phases, ready to take control in an emergency, eventually the car will run autonomously. But most people are reluctant to cede total vehicular control to a bot, so the Academy of Robotics built its own version of Mission Control: the Command Hub. The massive data center houses an array of supercomputers that receive and process every bit of information from the Kar-go’s internal and external sensors. Humans monitor sensor and diagnostic information, ready to take control of the vehicle and operate it remotely, if necessary.
Kar-go’s Future
I asked Sachiti about the company’s business model—is he planning to sell the vehicles outright, lease them individually or by the fleet, or offer delivery as a service? He said that at this point, the company is running initial trials with a few large companies that pay for deliveries. Depending on the results, Sachiti’s company will “follow the path of least resistance,” which could turn into a leasing strategy, or it may sell fleets to delivery companies. In the meantime, he’s looking for investors and talent, especially engineers who have significant experience in AI.
Sachiti closed the interview with this parting thought: “What makes Kar-go magical for me is that we applied artificial intelligence and robotics in a useful, but good way: the technology is there when it is needed and out of the way when it isn’t. Unlike many iterations of AI on the Internet today that want something from you or want to keep you in some app or drive you to make buying decisions, we don’t. Kar-go is a very complex machine performing a simple task—only when required. That is an AI-assisted future I want to live in.”