NORI L3 Review: The $1,688 Home Robot, Examined

What the NORI L3 Is, and Why Nori Robotics Also Calls It the A3

The NORI L3 is a bimanual mobile manipulator: two arms on a lifting central column, mounted on a wheeled base, sold for home use at $1,688. Nori Robotics is a San Francisco company founded in 2026 by Antonio Sitong Li. It lists five employees and appears in the Y Combinator Summer 2026 batch, where it describes the product as "a sub-$2000 bimanual mobile manipulator that can clean rooms, restock shelves, and fold clothes."

The name is genuinely inconsistent across sources, and buyers searching for the L3 should know why. Press coverage from Forbes and tech360 in August 2026 calls the machine the Nori L3. The company's own site at norirobotics.com heads the same product, price, and spec list as the NORI A3. The earlier product domain noril1.com now serves the same A3 page. Treat L3 and A3 as one machine until Nori publishes a naming note, and check the model designation on any order confirmation before paying.

The physical layout matters more than the label. Two arms sit either side of a vertical core that raises and lowers the head and arms together. That column is the reason the machine can work at floor height and at counter height without a leg system, and it is the single design decision that keeps this robot in the four-figure bracket.

Wheels, Not Legs: The Distinction That Defines the Price

The NORI L3 moves on wheels and does not walk on two legs, which is the main reason it costs $1,688 rather than five figures. Bipedal balance is the expensive part of a humanoid. It needs high-torque actuators, fast control loops, and enough structural margin to survive falls. A wheeled base removes all of that and replaces it with two drive wheels and a flat floor assumption.

The trade is concrete rather than philosophical. A wheeled robot cannot climb stairs, cannot step over a raised threshold, and gets stuck on deep pile rugs and cables. In a single-storey apartment with hard floors, none of that matters. In a two-storey house, the machine is confined to whichever floor you carry it to. Anyone shopping for a robot to work across a whole home should treat that as a purchase-blocking limit, not a minor caveat.

Calling a wheeled dual-arm platform a "humanoid" is a marketing convention rather than a technical claim, and it is common across this price tier. The useful category is mobile manipulator. Judge it against other mobile manipulators and against a good vacuum plus a human, not against the bipeds in our humanoid robot price comparison.

What Nori Robotics Publishes, and What It Does Not

Nori publishes six specification lines and no more, so several figures a buyer would normally check are unavailable. The company-published sheet covers the arms, lidar, cameras, audio, battery, and assembly location. Height, weight, reach, drive speed, compute, warranty terms, and any autonomy benchmark are absent from the vendor page.

Third-party numbers circulating for the L3 do not agree with each other, which is a reason to wait for Nori's own figures rather than to fill the gaps. Press reporting gives 19 degrees of freedom and a 55 kg vertical lift on the central column; the vendor page instead gives 7+1 DOF per arm and says nothing about lift capacity. The two DOF counts reconcile cleanly if the total is 16 arm joints plus a head pan-tilt pair plus the lift axis, but that is arithmetic, not a published breakdown. At least one robot directory lists a height of 145 cm, a weight of 40 kg, and a country of origin of China, all three of which contradict Nori's own material. Ignore directory listings for this machine.

The payload figure is the one to read closely. Nori states 1.5 kg per arm with the arm extended. A full one-litre carton of milk weighs about 1 kg, so a two-litre jug is already outside single-arm range. The 55 kg number reported in the press describes what the ball-screw column can raise vertically, not what the hands can carry, and conflating the two produces a badly wrong picture of what the robot can lift off a shelf.

What "Assembled in San Francisco" Means and What It Does Not

Assembled in San Francisco means final assembly happens in the United States; it does not mean the parts are American. Nori states the assembly location on its own site and makes no claim about component origin. Reporting on the launch notes that most internal components are likely sourced overseas, and the company's own research lineage supports that: the design descends from an open hardware project built around Feetech serial-bus servos, which are manufactured in China.

For a buyer this distinction changes three practical things. Lead times are exposed to component supply rather than to a US factory schedule. Tariff or import-rule changes on robot components can move the price of a later batch. And service depends on Nori holding spare servos in San Francisco, which the company does not currently publish anything about.

The upside is real and worth stating plainly. A US assembly line means a domestic point of contact, domestic shipping, and no customs clearance for US buyers. That is not true of the Chinese-built machines in our Chinese humanoid robot buying guide, where import duties and clearance are the buyer's problem and routinely add hundreds of dollars to a sub-$5,000 order.

How the NORI L3 Relates to the L1 and the L2

The NORI L3 is the third machine in a fast, cheap iteration cycle that began with a sub-$1,000 research prototype in May 2026. The lineage is documented well enough to trace. In May 2026 Antonio Li and two co-authors at Columbia University published a paper titled "Nori Bot: A Sub-$1,000 Floor-to-Counter Mobile Manipulator," describing a 17-DoF dual-arm mobile manipulator with a $947 bill of materials.

That paper is the most useful document about this product family and almost nobody covering the L3 has read it. It states the Nori Bot was built on top of XLeRobot, a $660 open-source dual-arm mobile robot, and that the main addition was a 600 mm Z-axis lift assembled from stock CNC parts: two HGR20 linear rails, four HGH20CA blocks, and an RM1605 ball screw driven by a Feetech STS3215 servo. It also names the failure mode the design was built to survive: stall-induced burn-out that destroys cheap Feetech servos. A robot in this price bracket lives or dies on that specific problem.

The commercial ladder that followed is short. The L1 was announced by the founder as a sub-$1,000 machine you could pre-order for summer 2026. The L2 was covered in July 2026 from $1,288 with remote operation and a skill marketplace, and its production run was limited to 100 units. The L3 arrives at $1,688 for fall 2026. Each generation adds capability and about $400. Three product launches in roughly four months is a hardware startup moving faster than most can support in the field, and support is the part that is not visible yet.

• Nori Bot (research, May 2026): 17 DoF, $947 in parts, built on XLeRobot, published as an arXiv paper
• NORI L1: announced as the best robot under $1,000, pre-order, summer 2026 delivery
• NORI L2: covered from $1,288, remote operation plus skill marketplace, production run capped at 100 units
• NORI L3 / A3: $1,688, fall 2026, first Nori units reported shipping from 21 July 2026

Teleoperation Versus Autonomy on the NORI L3

The NORI L3 is trained by a person driving it and then runs the recorded skill back, which is imitation learning from teleoperation rather than general autonomy. Nori supplies a laptop application called Nori Lab to train, operate, and manage the robot, and a Skill Marketplace where owners publish skills they have taught their own machine and install skills taught by others. The L2 generation was also described as supporting remote operation, so a person elsewhere can drive the robot directly.

The research paper makes the maturity level unusually legible. Data was collected by leader-follower teleoperation at 50 Hz, with one arm used as the leader while the robot copied it. The team then trained a single ACT policy on the right arm for one representative skill, picking a small DC motor off a desk, from 30 demonstrations. The paper reports that policy works with the object inside the trained spatial distribution, under neutral lighting, on a clean background. The authors also state directly that they had not yet logged a controlled multi-trial sweep with categorised failure modes.

That is the state of the art at this price, and it sets expectations correctly. A skill trained on your kitchen, in your lighting, with the mug in roughly the same place, is a reasonable thing to expect. A robot that finds the dishwasher on its own the first time is not. The Skill Marketplace inherits the same limit: a pouring skill trained in someone else's kitchen encodes their counter height, their bottle, and their lighting, and the paper explicitly notes it does not implement cross-unit pooling. Expect downloaded skills to need retraining rather than to work on arrival.

The Chore List, Assessed Against the Hardware

Nori lists six household tasks, and the hardware supports some of them far better than others. The published list covers helping in the kitchen, tidying, fetching from the fridge, loading dishes, folding clothes, and pouring ingredients. Sort them by how much force, precision, and variation each one needs.

Fetching a light item from a fridge and moving objects between fixed positions are the natural fit. The lift column gives floor-to-counter reach, the lidar handles room-scale navigation, and the task repeats in the same place every time. Folding clothes is the hardest item on the list for any robot at any price, because fabric has no fixed shape and the paper notes the current grip-force sensing under-resolves on compliant objects such as sponges and fabric. Loading dishes sits in between and depends heavily on your specific rack.

Anyone treating this as a general solution for household chores will be disappointed, and anyone treating it as a programmable platform that does three or four chores well in one room may not be. The distinction is worth more than any spec on the sheet.

NORI L3 Price Against the Rest of the Market

At $1,688 the NORI L3 undercuts the cheapest full-body humanoid this site tracks by roughly a factor of three. The comparison is not like for like, because the machines below it on capability are wheeled and the machines above it are bipedal research platforms with far higher build quality.

The buying detail that separates this from most launches is the payment structure. Nori states "$1688, full price, no deposit" for a fall 2026 delivery. That is not a refundable reservation of the kind used elsewhere in this market; it is payment in full ahead of shipment to a company with five employees and three product generations in four months. First units are reported to have shipped from 21 July 2026, so this is not a concept, but a paid pre-order with no deposit option puts the delivery risk entirely on the buyer. Weigh that as carefully as any specification, and check the refund terms in writing before ordering.

For context on the tiers above this one, see our cheapest robots to buy guide and the affordable options under $30k breakdown.

Who the NORI L3 Suits and Who Should Skip It

The NORI L3 suits people who want to teach a robot tasks and accept that teaching is the work, not a bug in the product. That includes robotics students, makers, developers building on top of a cheap manipulator, and early adopters in single-floor homes with hard floors and a tolerance for a machine that fails in front of guests.

Skip it if you want an appliance. Nobody should buy this expecting to unbox it and have laundry folded. Skip it if you live over more than one floor and need the robot to move between them. Skip it if you cannot afford to lose $1,688 to a delivery slip, because the company takes full payment in advance and publishes no warranty, repair, or spare-parts policy. Skip it if you need a support contract, since a five-person company cannot offer one credibly yet.

One buying criterion beats every spec here: ask whether you would still be glad you bought it if the Skill Marketplace stayed empty and you had to train every task yourself. If the answer is yes, the hardware is priced well for what it is. If the answer is no, you are buying the ecosystem, and the ecosystem does not exist yet.