The 2002 Roomba Cost $199.95 and Had No Map; Today's Humanoids Share a Charging Problem and a Lidar Supplier With the Robot Vacuum Industry but Little of Its Hardware

iRobot introduced the Roomba in the United States on 18 September 2002 at a suggested price of $199.95, and the machine had no idea where it was. It had a soft bumper, sensors that detected stairs, and a few movement patterns. Unitree lists its G1 humanoid at $13,500 before tax and shipping, and the G1 carries a depth camera, a 3D lidar and 23 joints.

A tidy story says the second machine grew out of the first. The documents say something narrower. Robot vacuums and humanoids face a shared list of problems, including knowing where you are, avoiding a fall and getting back to a charger, and in at least one case they buy from the same supplier. They share very little of their moving hardware. The humanoid side of the comparison, including the pilots at BMW and Haneda Airport, is covered in our piece on humanoids clocking in.

A robot built not to know where it was

Joe Jones, a named inventor on the Roomba patent, traced the machine in a 2020 essay to a 1989 prototype called Rug Warrior, which had bump sensors and a bottle-brush carpet sweeper. He calls it “the earliest conceptual ancestor of Roomba.” Its control scheme was what Jones describes as behavior-based programming. A behavior called Cruise drove forward and ignored every sensor. A second, Escape, watched the bumper switches and turned the robot away from whichever side hit. An arbiter, a small piece of code that settles conflicts, let Escape win.

A map was never on the table. Jones writes that a position-sensing system around the year 2000 would have added more than $1,000 to the Roomba’s cost. So the engineers made the robot work without knowing where it was.

The patent, US 6,883,201, was filed on 16 December 2002 and granted on 26 April 2005 to Jones, Newton Mack, David Nugent and Paul Sandin. It describes a “cliff detector” built from an emitter and a detector that sense a drop in the robot’s path, and a bumper that triggers a “bounce” mode, a random or weighted-random turn. iRobot’s launch release adds the rest of the picture: a spiral pattern until the bumper hits something, then wall following, covering each area roughly two to five times. A removable battery gave 1 to 1.5 hours. A separate “virtual wall” unit could block a doorway up to 15 feet wide. Covering each area several times stood in for the map.

The map arrived from a lab paper and a cheap laser

The mathematics behind mapping was older than the Roomba. Cadena and colleagues, writing in IEEE Transactions on Robotics in 2016, date the “classical age” of simultaneous localization and mapping (SLAM, the method by which a robot builds a map while working out its place on it) to 1986 through 2004. The sensor was the holdup.

In 2008, Kurt Konolige and four co-authors presented a spinning laser rangefinder at the IEEE robotics conference in Pasadena. They argued that the cost and power draw of existing laser scanners blocked affordable home robots, and reported a build cost under $30 for a unit with 3 cm accuracy out to 6 m, a 10 Hz scan and 1 degree resolution over 360 degrees. When SparkFun took apart a Neato XV-11 robot vacuum in November 2010, it reported a laser sensor with 1-degree planar scanning at 10 Hz, noted that a white paper claims such a sensor can be built for under $30, and added that the laser could then be bought only inside the robot. The XV-11 cost about $400. Its other parts were ordinary: optical switches for cliff detection, two 7.2-volt nickel-metal-hydride battery packs and a charging base with spring-loaded contacts.

iRobot’s own mapping came later. When it announced the Roomba 980 on 16 September 2015, it said it was using its vSLAM technology “for the first time in a consumer product.” The robot built a map as it cleaned, ran up to two hours and returned to recharge. It started at $899. Thirteen years separate the first Roomba from the first one that mapped.

What “largely solved” covers, and what it does not

Cadena and co-authors ask whether SLAM is solved and answer that the question only makes sense for a given robot, environment and performance target. They write that mapping a 2D indoor space with a robot carrying wheel encoders and a laser scanner “can be considered largely solved,” and that other combinations still need fundamental research. A vacuum on a flat floor is the solved case. By that framing, a walking robot with arms is a different combination.

The sensor suppliers show where the two industries touch. Hesai, a Chinese lidar maker, reported in its second-quarter 2026 results, dated 18 August 2026, that it shipped 142,371 robotics lidar units, up 193.4% from 48,531 a year earlier. Its automotive driver-assistance lidar shipped 485,904 units in the same quarter, so cars still dominate. Hesai said more than 50 embodied AI companies, including Unitree, had adopted its JT128 model. Gasgoo reported on 9 January 2026 that Hesai also disclosed ranking first in 2025 for 3D lidar shipped to lawn-mowing robots, and that Dreame, a large robot-vacuum brand in IDC’s trackers, was among the first to scale lidar in mowers.

The same supplier thus sells into a vacuum maker’s lawn-mowing robots and into humanoid makers’ bodies. That is a shared vendor, and none of the humanoid company pages used here describes its hardware as derived from a vacuum.

Docking and batteries: the same chore at a different scale

The 2002 Roomba ran 1 to 1.5 hours, and the 980 ran about two. The G1 lists about two hours, the same as the 980, with a quick-release 9,000 mAh battery.

Humanoid makers have their own answers to getting back to a charger. Boston Dynamics says its Atlas, which it lists with a four-hour battery life and 56 degrees of freedom, “autonomously navigates to a charging station and swaps out its own battery.” Figure’s third-generation robot, announced on 9 October 2025, has charging coils in its feet and charges at 2 kW by stepping onto a wireless stand. Returning to a charger is routine for a vacuum like the 980. Atlas and Figure 03 do it with a battery swap and a stand to step onto.

Where the hardware parts ways

A robot vacuum’s drive is simple. SparkFun’s Neato teardown found two drive wheels, each a one-piece motor and gearbox with a digital encoder. A humanoid has to hold its own weight on two legs. The G1 lists a maximum knee joint torque of 90 newton-metres.

The actuator ideas behind legged robots came from a separate line of work. Patrick Wensing and colleagues wrote in 2017, in a paper on MIT’s Cheetah quadruped, that conventional industrial actuators lack high torque density and handle dynamic physical contact poorly. They proposed a proprioceptive design, in which the joint itself senses and controls contact force and can be pushed back by an impact. Their subject was a four-legged robot.

Falling is the other split. A Roomba’s cliff sensor keeps it off the stairs. For a walking robot the fall is the hazard. Rodney Brooks, who lists the Roomba among five families of robots he helped build, wrote on 26 September 2025 that a falling legged robot carries a lot of energy in its legs and that a robot 50% larger needs about 3.4 times the energy. His advice was not to come within 3 metres of a full-size walking robot. He expects humanoids to get wheels for feet.

Hands are a third. Brooks cites a 1979 review counting about 17,000 low-threshold touch receptors in the skin of the human hand and argues that “Collecting just visual data is not collecting the right data.” Figure says each fingertip sensor on Figure 03 can detect forces as small as three grams. The difficulty of the hand is the subject of our piece on how robot hands learn and of the old puzzle in Moravec’s paradox.

Who owns what now

IDC counted 32.72 million cleaning robots shipped worldwide in 2025, up 20.1%, of which 24.12 million were smart vacuums. It also said iRobot had fallen out of the global top five. On 14 December 2025, iRobot announced that Picea, its secured lender and primary contract manufacturer, would acquire 100% of its equity through a court-supervised process. iRobot’s release says the company introduced the first Roomba robot vacuum in 2002.

Vacuum makers are adding arms. Roborock announced the Saros Z70 on 6 January 2025 with a foldable five-axis arm that lifts items under 300 grams, based on the manufacturer’s own testing. Ecovacs vice chairman Qian Cheng told Yicai in March 2025 that the company had invested in sensor suppliers, chip developers and robotic-arm makers, and that household humanoids would need more joints, more load capacity and coordinated control, according to the report. He gave no launch date. “We don’t want new products to be stitched from vacuum cleaners and robot arms,” he said.