The short answer: a multi-purpose home robot is three things: a body that can move around a home, arms and grippers that can handle objects, and software that decides what to do from what the robot’s cameras and sensors see. The body and arms decide what the robot could do. The software decides what it actually manages, and whether it can cope when things are not exactly as expected. For almost every robot on this site, the software is the least-known part.
The body: wheels or legs
Home robots get around in one of two ways.
- Wheeled robots roll on a base, with a torso and arms above it. Isaac 1, Futuring 2 and Bajie are examples. Wheels are simpler and more stable, but a wheeled robot cannot climb stairs and can struggle with narrow passages and high shelves.1
- Legged humanoids walk on two legs, in a roughly human shape. NEO and GR-3 are examples. A human shape suits a home built for humans, but walking robots have historically been far less reliable outside the lab.1
Neither is better in general. What matters is whether the robot can reach the places in your home where the work is.
Arms and hands
To appear on this site a robot must be able to pick up and move things. How it does that varies a great deal:
- Simple grippers: two fingers that open and close. Isaac 1 has two arms with a simple gripper on each.
- Dexterous hands: several jointed fingers, closer to a human hand. NEO and GR-3 have hands.
The type of hand does not by itself say what a robot can do: look at what it has actually been shown doing. How much weight each arm can lift and how far it can reach are listed on each robot’s page where the vendor has published them.
The software: scripted or learned
This is the most important difference between robots, and the hardest to see from outside.
A scripted robot replays a fixed sequence of movements. It works when everything is exactly where it expects, and fails when a shirt is crumpled differently or a cup is a few centimetres to the left.
A learned robot has been trained on many examples, often recordings of people guiding the robot through a task by remote control, and produces its movements from what it sees at the moment.2 In principle it can adapt: if a sock slips mid-fold, it notices and tries again. This site only lists robots with evidence of this kind of adaptation.
Some vendors name their approach, and those names are recorded on each robot’s page: for example 1X’s “Redwood” model for NEO, SwitchBot’s “Omni Sense VLA” for onero H1, or Sunday Robotics’ training of Memo on recordings made in hundreds of homes with a sensor glove. A “VLA”, one of the most common terms, is explained in VLAs and teleoperation, explained.
How a task gets started
Robots differ in how you ask them to do something: by speaking to them, by pressing a button in an app, by showing them, or on a schedule. This is an interface choice, not a measure of intelligence. A robot started from an app can be just as adaptive once it begins as one started by voice.
What the robot is allowed to try
A separate question is whether the maker lets the robot attempt a task it has not been checked on. Some vendors restrict their robot to a list of tasks they have tested, even where the underlying software could attempt more. NEO is one: 1X says it handles the chores it has been trained on, and a remote human operator guides it through the rest. For most robots the vendor has not said, and this site records that as unknown.
Remote human help
Some robots are partly operated by a person working for the vendor, who can take over remotely when the robot gets stuck or meets a task it cannot do. NEO and Isaac 1 both work this way, and their pages say so. This site does not treat that as disqualifying, but it always labels it, because it matters to a buyer.
Notes
-
The Ohio State University News (Ayonga Hereid), “Why household robot servants are so hard to build”, 2022-09-22. https://news.osu.edu/why-household-robot-servants-are-so-hard-to-build/ ↩ ↩2
-
Zhao, Kumar, Levine and Finn, “Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware”, arXiv, 2023-04-23. https://arxiv.org/abs/2304.13705 ↩
Last updated 2026-09-27.