We set ourselves a deliberately extreme target: a sidewalk delivery robot for $500. We missed it. What we got from trying is worth more than the target.
The problem is still open. Walk through a city that has delivery robots and you will see many of them parked on the sidewalk, waiting. The numbers say the same. Serve Robotics, the largest listed operator, has more than 2,000 robots deployed, and in Q3 2026 fewer than half of them made deliveries on an average day. Starship, the largest operator overall, announced in June that it is moving more than 1,200 robots off the U.S. campuses where it started and into grocery delivery. We read both as the same sign: nobody has yet found the place where a robot at the curb is clearly worth what it costs. A much cheaper robot is our bet on changing that.
The AI works, on cheap hardware. The navigation policy is trained entirely in our own simulator and drives the real robot on real sidewalks without fine-tuning. It runs on board, on the NPU of a single-board computer that costs $300 at most, at 10 Hz with 200 ms latency from cameras to wheels. At walking speed the robot moves about 30 cm in that time, which is short enough for a sidewalk.
The device costs $1,000, not $500. We built a $500 robot and it drove, but that left no money for a good wheel base or motor controllers and the device was not reliable. With hub-motor wheels, a waterproof 2D lidar and a proper controller, the parts add up to about $1,000. Chip prices are moving with the shortage, so the exact figure will shift, but we expect the final product to stay in that range. That is still a fraction of what sidewalk robots cost today.
The rebuilt prototype is due in January 2027. Same AI, same teleoperation, new wheel base. We will post video when it drives.