How Kodiak Teaches Autonomous Vehicles to Read the Road
|
Not all obstacles are created equal. Kodiak AI is pioneering novel ways to address distinct types of obstacles we encounter on the road.
In the world of autonomous driving, some objects are easier to negotiate than others. Detecting and avoiding large objects, for example, represents a straightforward task. But detecting and responding to obstacles that lurk at or below ground level – potholes, puddles and ditches, for example – presents a difficult but important challenge for driverless vehicles.
In the realm of Physical AI, Kodiak calls these and other types of voids in terrain “negative obstacles.”
Negative obstacles are the absence, rather than the presence, of mass in a particular space. While they do not present a collision risk per se, any motorist who has inadvertently struck a pothole at high speed recognizes they hold the potential to topple cargo, flatten tires and cause structural damage to a vehicle.
Kodiak’s AI-powered autonomous driving solution is designed to tackle tough driving jobs. This includes in long-haul trucking applications on highways, along bumpy dirt roads in industrial settings, and in rugged environments for defense applications, which often include driving in places where roads do not exist.
To address the challenge that negative obstacles present and ensure seamless operation, Kodiak has developed an end-to-end “Traversability Framework” that identifies negative obstacles and assesses their depth and other physical dimensions.
Classifying a negative obstacle is only the first step. The framework utilizes continuous behavioral decision-making in helping the Kodiak Driver determine whether it should drive through, nudge around or straddle a negative obstacle. It also helps define the proper speed for those maneuvers or decide if stopping for further evaluation is warranted.
Kodiak’s Traversability Framework distinguishes between harmless water reflections and more problematic, puddle-filled potholes through an AI-based analysis of sensor data. The framework integrates across our autonomy stack, ensuring negative obstacle planning complements avoidance strategies for more traditional obstacles.
Some AV programs have optimized their autonomy stacks for detecting positive obstacles, and they address negative obstacles by stopping or avoiding them. This can lead to vehicle immobilization, which is problematic when operating on surface streets and highways.
Such an approach is also untenable in industrial and defense environments, where unforgiving roads with potholes, ruts, dips and ditches are as much the norm as the exception. The framework has been a critical enabler for our deployments in these verticals. By integrating negative obstacle considerations into our planning, we have taught our trucks to negotiate terrain in ways standard binary planners cannot.
Our unique approach gives us an edge: our system makes nuanced, human-like decisions and maintains the highest-possible speed through continuous trajectory profiling and planning, rather than crawling at pre-set, low speeds or stopping when negative obstacles are detected. It allows us to handle scenarios and terrain other systems may deem impassable, ensuring mission continuation and success.
The Traversability Framework even accounts for the requirements of specific vehicle platforms: a truck hauling double trailers has very different needs from a Kodiak Driver-equipped Ford F-150 designed for defense applications or the Textron System’s RIPSAW M3 treaded vehicle upfitted with our technology. Kodiak uses geometry and visual appearance as inputs to map the terrain against the physical anatomy of the vehicle.
That foundation has informed our ongoing autonomous freight-hauling work on highways across the southern United States, and it has girded our work with the Pentagon and our commercial deployment in the Permian Basin, where our customer, Atlas Energy Solutions, uses autonomous trucks to haul frac sand. We’ve seen and handled thousands of potholes there. This is a prime example of how learnings from one application accelerate progress across others.
By turning negative obstacles from hidden threats into a well-managed engineering framework, Kodiak has shown the mettle for handling uncertainty in the roads ahead. That’s the difference between autonomy that merely moves, and autonomy that adapts, survives and thrives in the physical world.
Sandeep Reddy Baddam is a motion planning engineer who has led Kodiak’s work on negative obstacles. Previously, he served as a research assistant on the DARPA RACER (Robotic Autonomy In Complex Environments with Resiliency) program at the University of Washington and, separately, worked on autonomous technology for NASA’s Lunar Rover while at the Bosch Center for Artificial Intelligence. He can be reached at linkedin.com/in/sbaddam
