Methods

« Back to Homepage

Tweet

The Road to Level-5 Autonomous Vehicles
Design Trends and the Transformation of Everything

Read the full Design Trends eZine:

---

By Sudha Jamthe for Mouser Electronics

Autonomous driving technology is advancing rapidly, with all the major players testing level-4 vehicles. Most expect to test level-5 vehicles on public roads within the next few years, and this activity has spawned its own ecosystem of specialized sensors, power systems, data processing, and software.

Most autonomous vehicle (AV) players today say that they’re testing vehicles that are at level 4 (based on the Society of Automotive Engineers zero-to-five scale). That means these vehicles are technically able to operate without human input, but humans can take over if the AV gets into trouble. True, these vehicles typically run in limited geographies restricted by geofencing, but this feat is still an impressive accomplishment and the cars keep getting better. All the players in this space are striving for fully autonomous level-5 vehicles. Waymo, Google’s AV project, is already preparing to test level-4 vehicles without a human backup driver on public roads in California.

Essential Capabilities

To understand what has been accomplished and what still needs to be done, consider these four essential capabilities a self-driving car must have to function without human input:

Localization

An AV must always know where it is relative to its known world. It must know its starting point, whether to turn, how to navigate around an obstacle, and where to detour when a road is closed. If the vehicle shuts down for a time, it must be able to put its current location on the “map.” Notice that I say it must localize itself in its known world: This is because AVs are not likely to know the entire world. Level-4 AV testing typically begins with a restricted area. As the vehicle gets better at negotiating this area, testers expand its routes. When AVs become commercially viable, many will function in limited ranges and routes. Regardless of the size and shape of an AV’s world, the vehicle must always be able to localize itself at all times.

Perception

An AV must be able to perceive the physical world around it, including recognizing roads, lanes, traffic signs and lights, other vehicles, pedestrians, obstacles, and all the things humans see and process as we drive. The most common AV sensor technologies include video cameras, stereo video, radar, and light detection and ranging (LIDAR)—a pulsed laser technology used for object detection and avoidance. Currently, no standard approach exists for the perception layer. Many players are testing various combinations of sensor arrays, and cost is obviously a factor. Whatever sensor array an AV has, it must be able to stitch together all the data from all those sensors to build a detailed and complete local map of its surroundings, which updates continuously in real time as the vehicle moves.

Path Planning

An AV must be able to make every navigational and operational decision in every moment as it’s traveling. This decision-making goes beyond getting from point A to point B in its known world. The vehicle must also make decisions about what to do next if it stops, when it should speed up or slow down or change lanes, how to execute turns, what to do if it encounters an obstacle, and what to do if it encounters something it has never experienced before. Of course, it must be able to decide when to leave point A and what to do when it reaches point B.

Control

An AV must be able to reliably and accurately operate the proper sequence of driver controls to empower it to perform essential operations. If the AV decides it needs to go around an obstacle, and doing so requires a lane change, it must use its sensor array to ensure the path is clear to perform the lane change without violating traffic controls for that stretch of the roadway; it must then execute the proper speed adjustments, signals, and steering to perform the action.

All four of these capabilities are essential for an AV to function at all levels. Commercially viable level-5 AVs must exhibit these capabilities with considerable accuracy and reliability and be able to react to all road situations independently without human intervention. Achieving this level of reliability depends on another critical piece in an AV system: The artificial intelligence (AI) platform, with its deep machine learning (ML) and inference modeling, as well as its training to drive.

What’s Happening Today

Much of the level-4 testing happening today is really training the system to react properly in every possible driving scenario. Typically, this work starts small, with the test vehicle traveling a narrowly prescribed route over and over. Through this process, the AV is being taught the basics through simulation. Every time it travels a test route, all its operations are logged. If the AV gives control to a human driver, it’s called a disengagement, and that human intervention becomes part of the log. Every AV disengagement is reported to the DMV by regulation. When the car returns to the shop, engineers tweak the vehicle’s AI brain, using what it learned on that test run. They then create a new inference model for the AV to use when the vehicle goes out again. This process is complex deep learning that involves processing large volumes of | 28 unstructured data and ultimately being able to make decisions in real time. Human interventions at this stage are a critical part of the training process.

As the AV masters a route, the testers expand the vehicle’s range to take in new road situations and geographies. That new geography may cover a community and then expand to an adjacent community. Waymo has been training level-4 AVs in Phoenix, Arizona, and several other cities. It’s worth noting that many of these testing programs cover routes and regions that would potentially support an AV-based business model, such as Mobility as a Service. Major players are being strategic about where and how they test their vehicles.

The driving knowledge the AI system gains is cumulative and transferable to other vehicles. At the level-4 testing stage, driving knowledge enhancements may look like beta code updates. When advanced AVs become commercially viable, they will receive regular updates of accumulated knowledge. The key is that the more miles AVs log, the better and more reliable they become because they’re continuously learning systems.

What’s Ahead

In the coming year, we will see more about level-5 AV road tests, and AV developers will continue to address the engineering challenges that remain. For example, current AVs do not fit into the social fabric of the road. If you honk at an AV, it will not know what to do. It cannot communicate with pedestrians or other drivers. Recent industry discussions have focused on establishing communications standards to avoid the chaotic situation of every AV manufacturer having its own communication protocols.

Level-5 AVs will ultimately have no need for driver controls, and there will be no reason for everyone to face forward. AVs will become high-bandwidth devices capable of streaming media for passenger consumption, and they will have to do so without disrupting the control systems and connectivity associated with a vehicle’s operation. These design parameters call for a complete redesign of car interiors.

It’s a tremendously exciting time for engineers in the automotive space. AVs are poised to transform the auto industry through sensory and control systems and vehicle ergonomics up to a greater diversity of purpose-built vehicles and an expanding role for AVs in the continued emergence of all-electric vehicles. It may all be happening faster than we think.

Sudha Jamthe is CEO IoTDisruptions.com and DriverlessWorldSchool.com, as well as a recognized technology futurist. She has more than 20 years of experience with companies like eBay, PayPal, and GTE. Sudha is also the author of 2030: The Driverless World and three books on the Internet of Things. She teaches courses about the business of IoT and autonomous vehicles at Stanford Continuing Studies and DriverlessWorldSchool.com, is chair of the strategic advisory board for Barcelona Technology School, and is an ambassador for FundingBox Impact Connected Cars