Robo Rigs - A Heavy Commitment
Autonomous semis are closer to reality than you might think.
Innovating companies like Kodiak Robotics, Aurora, Waabi & a few others have been investing billions of dollars in level 4 r&d to develop cutting edge sensors (reminder: SAE' level 4 is at the ML degree vs. level 5 being fully cognitive) & some are already hitting the roads in the southern U.S.
The consensus is that there will be a massive deployment of autonomous Class 8 trucks within a decade, once each company aligns with their respective manufacturing teams. With a general focus on high-way driving, these trucks will also have to face FMCSA scrutiny before they get the green light & be officially welcomed as "Intelligent Rigs" (i'm not sure if the DOT will have to approve the tone of voice these ghost driver haulers will use when they get pulled over).
Every AV startup has its own unique advantage, engineering strategies & challenges.
The cost & difficulty of continuously seeking fresh real-world data, as well as inaccurate synthetic information that doesn't seize all contextual nuances are well known drawbacks in the field.
Some companies like Torc Robotics have prior experience building robots for DARPA & have partnered with manufacturing giants like Daimler (giving them a lead in CAN bus integration & building on existing safety/redundancy systems).
However, every autonomous trucking system has its own proprietary edge in terms of code stack, lidar, ultrasonics, scalability etc. combined with their hardware & integration of the NVIDIA DRIVE platform.
Kodiak Robotics was one of the first companies to deploy their driverless operations. Their approach & focus on modular hardware, fast retrofitting & exposing their stack to hybrid testing (training in various simulations e.g. when there is a breakdown or adverse driving conditions) on leased land, led them to obtaining a contract hauling frac sand. They've also secured hauling commitments with companies like Ikea and Werner.
Kodiak seized the opportunity to commercially haul proppant in the Permian Basin. Working closely with Kodiak gives Atlas Energy an edge for finding solutions (like customizing their trailers & improving logistics operations).
Those are early signs of progress, but that doesn't mean that 5 million+ truck drivers will loose their jobs in the near future. The adoption will be gradual and likely require a learning curve for operations managers (& a vote of confidence from confused traffic fighters).
You might recall a few years ago Google getting rid of Boston Dynamics and selling it to Soft Bank (not even 4 years after acquiring it) - Why? To put it simply, very few engineering teams in the world have the commitment to deal with the overwhelming and nuanced discipline of biological/natural mimicking (and implement that into hardware). The leaders of this field will be those that can strategically balance safety, compute efficiency & cost.
Pittsburgh based Aurora Innovation is another front runner in the autonomous engineering space. They put emphasis on hiring talent, acquiring startups like OURS (2021) & focusing to develop robust proprietary sensor fusion systems. The company's background is slightly complex (their stack originated under the Otto brand, then transitioned to Uber ATG, eventually leading to the creation of Aurora).
Nevertheless they have a strong team of engineers, have shown commitment for transparency, and have already secured contracts with companies (like Fedex).
Aurora's motion planning & perception systems, as well as a pledge for transparency with their Safety Case Framework positions them strongly for gaining community trust.
Waabi is another autonomous trucking company based out of Toronto, ON, Canada. Founded in 2021, their strategic approach in prioritizing computer vision first allowed them to build a unique generative AI autonomous stack.
Most AV companies took time in finding a way to pivot from traditional lidar scanning methods, data training (with techniques like NeRF/3DGS) & feed-forwarding. The challenge of slow pre-scene optimization & computational inefficiency have been haunting autonomous engineering teams forever.
Waabi decided to tackle this challenge head-on by developing its own 3D reconstruction configuration (G3R) & kept improving the framework like adding real time weather & dynamic simulators. This gave their AI a speed advantage in terms of reconstructing 3D environments.
They also created an advisory board of experienced drivers called 'Million Mile Driver', designed for truckers with over 1 million miles of safe driving to give feedback on r&d.
'Asymetric Self-Play' is a learning setup Waabi designed to train the company's AI, where a "teacher-student" framework, through iteration - allows the system to scale its capabilities & adaptation in complex scenarios.
A scalable proficiency to reconstruct quality scenes from lidar input is the cornerstone in autonomous trucking development. Waabi positioned itself as not only a computationally efficient pioneer in the space, but also cost effective.