Amazon’s autonomous vehicle subsidiary, Zoox, has officially launched its paid commercial robotaxi service in Las Vegas. The launch follows a historic Part 555 exemption from the National Highway Traffic Safety Administration (NHTSA), making Zoox the first company permitted to commercially operate a purpose-built passenger vehicle designed entirely without traditional human controls like steering wheels or pedals. Shaped like a compact, futuristic toaster, the vehicle represents a radical departure from conventional transportation. Rather than retrofitting an existing passenger car (as seen in Waymo’s Jaguar I-PACE fleets), Zoox built its vehicle from the ground up to be self-driving, lending to its unique design and features.
The physical architecture of the vehicle resolves several key challenges of urban driving. Because it features a symmetrical, bidirectional design with no dedicated front or rear, the car drives equally well in either direction. This eliminates the need to execute multi-point turnarounds on congested streets. Furthermore, independent four-wheel steering allows the vehicle to maneuver smoothly, execute zero-radius turns, and crabwalk diagonally into tight curbside pickup spots or narrow parallel parking spaces.
Now let’s look at the tech behind the car. Zoox relies on a redundant perception suite mounted on four floating sensor pods at the corners of the roofline. Each pod provides an overlapping 270-degree view, giving the vehicle an unobstructed 360-degree field of vision. High-resolution optical RGB cameras handle color, texture, traffic signals, and object classification. Alongside standard cameras, Long-Wave Infrared (LWIR) thermal sensors detect heat signatures, enabling the system to pinpoint pedestrians, cyclists, and animals in complete darkness, heavy fog, or intense sun glare. Rotating LiDAR sensors emit laser pulses to construct millimeter-accurate 3D point clouds of surrounding geometry, while radar arrays measure object velocity through dust, heat haze, or rain. To prevent environmental buildup from blinding the sensors, automated high-pressure air blasts and fluid sprayers clean optics on the fly.
Processing this continuous influx of data requires a localized supercomputing architecture housed in a liquid-cooled floor enclosure. Onboard NVIDIA GPU platforms run custom multimodal transformer models that fuse camera, LiDAR, and radar data directly in memory. Operating on a tight 100-millisecond execution loop, the AI recalculates its environment ten times per second, predicting traffic trajectories up to eight seconds into the future while evaluating potential driving paths. All critical perception and motion-planning calculations occur locally onboard, ensuring the car continues to drive safely even if cellular connection drops.
Now the most interesting thing is the mapping software. Zoox cannot just input or use Google or Apple Maps; rather, when they pick a city, they have human drivers drive some SUVs to map the area, especially where the traffic lights are, and create a 3D map. Of course, this map is going to capture things like people and cars, and using AI, they use it to take them out of the photo so they can have a clean map of the road. This takes terabytes of data, and so Amazon Web Services data centers help execute the brunt of the work. Raw, uncompressed sensor logs (LiDAR point clouds, radar scans, and high-res video) collected by mapping vehicles are uploaded directly to massive Amazon S3 Storage buckets while NVIDIA GPUs process this data while advanced neural networks (3D CNNs & Transformers) essentially edit the people out, much like a much more advanced version of Magic Eraser for cars. Then, with 5G, the maps are sent out to the cars.
Even if GPS fails, the car will still know where it is because of this advanced map, and the car’s onboard (local) AI will check with the LiDAR sensors and the onboard map to know its location down to the exact centimeter as well as the optical tracking. The four sensor pods run high-resolution cameras that track thousands of visual landmarks every second, such as building facades, bridge supports, and tree trunks. By measuring how fast these visual features move across camera pixels as the car rolls forward, the AI calculates exactly how far and fast the vehicle has moved.
When the vehicle encounters an ambiguous real-world situation (such as an unmapped construction zone or hand signals from emergency personnel), it connects via 5G to a TeleGuidance human operator at a Zoox Command Center. Rather than remotely steering the vehicle with a joystick, the human operator provides high-level semantic guidance or draws a suggested path, which the onboard AI then executes independently and safely.
God bless and Tech Talk To You Later!
Leave a Reply