Has Humble Robotics cracked the code on autonomous trucking?

The setup is an intuition that both speakers once held and that turned out to be wrong. Highways look easier than San Francisco: you mostly go straight, and less happens. That is why autonomous trucking outside defense began in 2016 with Starsky Robotics and Otto, and why Otto’s hundred-mile run with nobody in the front seat made the problem look nearly solved. What went wrong is not that highways have more edge cases but that they have different ones. They are rare, so the system gets few chances to encounter and learn them, and they are far less forgiving when they arrive, because the vehicle may weigh 80,000 pounds at full gross weight and its stopping distance is long. Cohen jokes that a truck with no perception system at all could probably hold a lane at 55 miles an hour for 200 miles; the difficulty is entirely in the uncommon event. Underneath that sits the asymmetry that matters most. A confused robotaxi can stop, which is annoying and occasionally viral but usually safe. A truck that stops on a highway, particularly past a bend or over a crest, has itself become the hazard. The graceful fallback that makes passenger autonomy tractable simply does not exist for trucking, and Cohen thinks that is much of why the timeline slipped. Ten years on there have been driverless runs, but no regular driverless service, and even the word driverless is contested: he notes that Aurora’s driverless runs sometimes carry a safety observer, citing the company’s public writings.

What has changed since 2016 is the brain, and not much else. Early stacks were hand-coded to the point of comparing pixel colors to find a lane line, and any understanding of the world had to be labeled in by people drawing boxes around objects. Today an off-the-shelf, sometimes open-source vision language model arrives already carrying a usable interpretation of a scene, including the sort of judgement that used to require explicit engineering: that a traffic cone sitting on the back of a pickup is being carried rather than marking a work zone. Cohen spent most of his career LiDAR-first and has flipped to camera-first at Humble because of that shift. He will not take the sensor debate as a matter of principle, though. For an 80,000-pound vehicle with very little margin for error he wants camera, LiDAR and radar all present, with one of them doing the heavy lifting depending on the task, since LiDAR cannot read a traffic light’s color, sees better at night, and radar can see through some weather. He accepts that a human drives with what amounts to two cameras and that trucks may eventually go vision-only, but treats that as an end state rather than as today’s engineering choice. Meanwhile the unglamorous parts, validation, safety engineering and hardware robustness, have matured steadily and are reused. So starting fresh means a new approach to the brain layered on well-understood practice for everything else.

The vehicle follows from asking what the simplest possible freight mover would be if no driver were ever involved: a box on wheels, a platform that carries a container. Humble’s truck is a cabless electric Class A vehicle that combines tractor and trailer into a single platform, and the justification is both technical and economic. Technically, a smart tractor pulling a conventional trailer cannot see directly behind itself, which rules out handling someone driving into the back of it and rules out backing into a dock, and produces oddities like the legal requirement to deploy warning triangles behind a stopped truck, something a driverless tractor cannot do; Cohen says Aurora, and he thinks Waymo when it was working on trucking, proposed high-mounted lights instead. Economically, deleting the cab removes both cost and weight. He states the counter-pressure himself: tractors and trailers are owned separately for good reasons, trailers being cheap and tractors expensive, and they get detached and sent in different directions, so a combined platform cuts against how the market is organised. He calls this a lot of challenges and does not describe how it is resolved.

Economics is where the episode’s own question actually lands. Freight is a business-to-business market with no cool factor to sell, so unlike Waymo, which Kann says is currently getting away with inferior unit economics because riders will pay more for the novelty, an autonomous truck has to save money to be adopted at scale. The headroom is the driver’s wage, roughly a dollar or a little more out of a fully loaded long-haul cost of about $2.30 to $2.40 a mile on the industry guide Cohen cites, and he prefaces the figures with a request not to be quoted on them. Against that sit the sensors, remote assistance, and the fact that retrofits use expensive high-end sleeper tractors because those have room for the equipment, which makes an already costly vehicle much costlier until scale arrives. Cohen’s response is to attack the vehicle cost rather than accept it, which is the actual argument for cablessness. He makes a parallel structural objection to the dominant long-haul model: hub-to-hub only works if the hubs exist, someone has to build them, and the short moves into and out of them are friction competing against simply hauling the load door to door. Humble instead targets short-haul and drayage, which is also where he thinks electrification genuinely fits, since long haul demands a very large and heavy battery and a great deal of charging power, a problem he says the industry still has to reckon with.