Madic Robotics co-founder explains vision-only home robot approach and path to multi-task humanoid

Summary

Jordy, co-founder and president of Madic Robotics, makes a case that floor-cleaning robots are not a dead-end product category but the foundation of a longer platform play toward full home humanoids. The company's current robot earned a rare 10-out-of-10 from Wired, and at roughly 1,000 units shipped, Madic is already the second-largest American consumer robotics company by volume — a figure that says as much about how thin the field is as it does about Madic's traction.

Vision-only, edge-first

Madic's core technical bet, made when the team left Nest and Google in 2017, is that indoor environments were built for human visual perception and therefore robots navigating them need the same system. The company runs everything on-device, reasoning that low latency is non-negotiable in dynamic home environments. The robot builds a photorealistic voxel map — each 1cm × 1cm unit tagged to an object class like chair leg, human leg, or child — using a combination of classical SLAM and neural networks rather than a single large model. Vision-language models like DINOv2 and CLIP extract semantic embeddings that let the robot understand natural-language commands such as "clean by the bookcase in the living room."

Compute costs have been deliberately constrained. The company has spent under $1 million on Nvidia compute in total, achieved partly by encoding physical priors directly into the system — if the robot already understands gravity and rigid-body physics, it does not need 26,000 iterations to learn that objects topple. Simulation built in Unreal Engine gets models to roughly 80% performance; the final 20% always requires real-world data.

The platform roadmap

The floor-cleaning form factor is framed as stage one of a child-development analogy: the robot first learns to navigate and map (years 0–5), then graduates to picking up unbreakable objects and simple organization tasks (years 5–10), before eventually becoming a full home robot. Jordy says demos already show the current perception and mapping stack running correctly when the sensor crown is elevated to six feet — meaning the same "brain" can slot into a humanoid chassis without retraining. Parents in the 200 homes he has visited consistently ask for a taller cleaning robot, and adult children with elderly parents in other states are requesting time-lapse welfare checks, pointing to two concrete near-term expansion vectors.

Privacy is handled through on-device processing and an opt-in model. Users who want to help improve the product can hit a record button in the app; Jordy says thousands of hours of labeled data have already been uploaded voluntarily, without any automated collection.

Humanoid market skepticism

On the broader humanoid wave, the arithmetic is sobering: iRobot has shipped 50 million units, Amazon Robotics 750,000, and Boston Dynamics roughly 1,500 across its entire history. Jordy's view is that purpose-built robots will reach consumers before general-purpose humanoids do, following the same arc as mobile computing — cell phones to PDAs to iPods to BlackBerrys before the iPhone combined them.

The accuracy problem is the sharper constraint. When AI gets 90% of a coding task right, users are impressed and handle the rest themselves. When a robot fails 1% of the time at a trivial task like setting a table, users consider it broken. The bar rises as the task becomes more mundane, not less. A humanoid priced at $40,000 — comparable to Unitree's current ask — faces expectations that are nearly impossible to meet at this stage of the technology.

The price ceiling matters structurally: no consumer electronics device above roughly $2,000 has achieved ubiquitous adoption. Cars cleared that bar only because their utility is unambiguous and they have had a century to prove it. Madic's implicit bet is to build trust incrementally through a sub-$2,000 product that works, then earn the right to charge more for additional tasks — rather than asking consumers to take a $10,000–$40,000 leap on unproven hardware.