이미지: X — 뉴스 앰프 영상 갈무리
Summary
- Matic's home cleaning robot builds a 3D map of the house using only five RGB-IR cameras and an NVIDIA Jetson Orin — no lidar — and updates that map in real time while cleaning
- Karpathy's vision-only argument wasn't about cost savings but "roads are designed for human eyes"; Tesla compensated with data for the sensors it removed from the product, but only at the training stage
- As lidar prices dropped from roughly $80,000 in 2016 to $200 in 2026, the cost argument evaporated — the remaining question is what more can be done with cameras alone
- 카메라 구성
- RGB-IR 카메라 5대, 라이다 미탑재
- 온보드 프로세서
- 엔비디아 젯슨 오린(가용 메모리 4GB)
- 지도 방식
- 1cm³ 복셀 기반 포토리얼리스틱 3D 지도, 절대 좌표
- 라이다 단가 변화
- 2016~2017년 약 8만 달러 → 2026년 200달러(허사이 ATX)
- 창업자
- 나브니트 달랄(2005년 HOG 논문 제1저자)·메훌 나리야왈라, 구글 네스트 출신
- 제품 가격
- 1,245달러 — 9월 9일 1,495달러로 인상 예고(부품 관세)
- 관련 업데이트
- 2026년 8월 13일, 손으로 가리키면 그 자리로 이동하는 제스처+음성 명령 공개
Cleaning a home without lidar
Robot vacuums are usually associated with a lidar sensor sweeping the ceiling, but Matic's home cleaning robot dispenses with that sensor entirely. Instead, it mounts five RGB-IR cameras across its body and reads the layout of the home purely from the footage they capture. The computation is handled directly inside the robot by NVIDIA's onboard computer, the Jetson Orin. Rather than sending video to the cloud for processing, the robot completes all its decision-making on the spot. It has only 4GB of memory to work with.
The camera footage collected this way isn't limited to simple obstacle-avoidance data. The robot runs a neural network that converts images into 1-cubic-centimeter voxels, reconstructing the home's structure into a photorealistic 3D map, and continues updating that map even while cleaning. If furniture moves or new objects are placed, the map updates accordingly.
Notably, this map is built using absolute coordinates rather than coordinates relative to the charging dock. The dock is treated as just one point on the map, so if a user picks up the robot and places it in a different room, it can relocate itself. The crown (top) and rear cameras reportedly carry infrared LEDs invisible to the human eye, allowing the robot to navigate the same way even in a dark house.
A digital twin of your home, built from cameras
| Component | Matic robot | Reference: lidar-based robots |
|---|---|---|
| Distance/structure sensing | 5 RGB-IR cameras | Rotating lidar sensor |
| Compute location | Onboard (NVIDIA Jetson Orin) | Onboard or cloud |
| Map updates | Real-time during cleaning | Periodic rescans |
| Coordinate reference | Absolute (dock is just one point on the map) | Typically relative to dock |
| Additional information | Includes visual data such as color and texture | Primarily distance values |
As the table shows, Matic's approach goes beyond simple obstacle avoidance, focusing instead on translating the home's structure itself into a visually rich digital map. The more refined this map becomes, the more the robot can do.
Indeed, on August 13, Matic unveiled another feature built on this 3D map. When a user says "Hey Matic, clean here" while pointing at a specific spot with their hand, the robot recognizes the direction of the arm, projects that direction onto its stored 3D map, and moves to the corresponding location. The camera-generated map has become the foundation not just for navigation but for understanding voice and gesture commands as well.
The day "lidar is a fool's errand" hit the stage
This design recalls Tesla's autonomous driving strategy from 2019. On April 22 of that year, at its "Autonomy Day" demo event, Tesla declared it would achieve autonomous driving using only cameras, without lidar. On stage, Elon Musk flatly stated, "Lidar is a fool's errand," adding that "anyone relying on lidar is doomed." The idea was to bet on AI that interprets camera footage the way human eyes do, rather than relying on expensive lidar sensors.
Lidar excels at precise distance measurement, but components are costly, and it can malfunction on certain materials like glass or dark-colored surfaces. Camera-based recognition, by contrast, is cheaper and can also capture visual information such as color and texture. However, calculating accurate distance and depth from cameras alone requires a far more sophisticated AI model. This is why the autonomous driving industry has long been split between lidar advocates and vision-only advocates.
At the time, Musk's remarks were mostly interpreted as being about cost. That's understandable, given that a single lidar unit once cost more than an entire car. But the argument presented by the person who unveiled the neural network component on the same stage was somewhat different.
Karpathy's argument wasn't about cost
Andrej Karpathy, who joined the stage alongside Musk, had been Tesla's head of AI since 2017. His argument was this: the entire road network is designed on the premise of human vision. Signs, lane markings, traffic lights, even hand signals at construction zones—all of it is information meant to be read by eyes. Lidar's point clouds can tell you "something is there," but they can't read what a sign says. So whether or not a car has lidar, the vision problem still has to be solved—meaning it makes more sense to pour resources into solving vision completely, rather than splitting effort across sensor types.
This logic became most clearly visible in 2021. That May, Tesla removed radar from its North American Model 3 and Model Y vehicles. Three weeks later, at the CVPR computer vision conference, Karpathy explained why. When radar and cameras produce conflicting readings, blending the two signals doesn't improve accuracy—it adds noise. Failures kept recurring: radar losing track when the car ahead braked hard, mistaking overpasses for stationary vehicles and triggering phantom braking, or flagging manhole covers as stationary objects. He summarized it as: "Everything needed for driving is contained within vision."
What stands out is what came next. Tesla removed the sensor only from the product, not from the development process. In the offline auto-labeling stage that generates training data ground truth, the company drew on additional sensors including radar, plus the advantage of "already knowing future frames"—switching to cameras only at inference time. Using shadow mode, which quietly runs inference on customer vehicles, the company set 221 failure triggers to identify difficult scenes, and cycling through this loop seven times over four months yielded 1 million hard-case clips and 6 billion labels. Vision-only was never about "removing sensors to cut costs"—it was a strategy of buying back what was removed, using data and compute instead.
Karpathy left Tesla in July 2022, spent time at OpenAI, and now runs the education startup Eureka Labs. In an October 2025 podcast interview, he said this about autonomous driving: it's still "not finished," and what remains is a "march of nines." Getting from 99% to 99.9% takes about as much effort as it took to build the first 99%, and the next nine is just as demanding. While acknowledging that Waymo's approach isn't yet economically viable, he maintained that Tesla's approach is more scalable. One line from him is still frequently cited: "Waymo has a hardware problem, and Tesla has a software problem."
Seven years later: lidar got cheap, and vision still isn't finished
The problem is that the premise "lidar is expensive" has since collapsed entirely.
| Timeframe | Product | Unit price |
|---|---|---|
| 2016–2017 | Velodyne 64-channel mechanical | ~$80,000 |
| 2020 | RoboSense RS-LiDAR-M1 (solid-state) | $1,898 |
| 2022 | Hesai ATX-class | $500 |
| 2026 | Hesai ATX improved | $200 |
Mass production by Chinese manufacturers pushed the curve down. That's a roughly 400-fold drop over a decade, and once prices fell below $200, lidar began to be treated as standard equipment rather than a premium option. Hesai's order backlog has surpassed 4 million units, and MicroVision has put sub-$100 pricing on its roadmap. The cost logic that underpinned Musk's 2019 statement has effectively vanished.
That doesn't mean the vision camp has won, either. To be precise, the matter is still undecided.
| As of early August 2026 | Waymo | Tesla Robotaxi |
|---|---|---|
| Sensor configuration | Lidar, radar, cameras | Cameras |
| Paid rides | ~500,000/week | Undisclosed |
| Driverless vehicles | ~3,500 | 20–40, in Texas |
| Cities of operation | 11 | 7 (safety monitor onboard in California) |
| Autonomous miles | ~4 million/week | Undisclosed |
In terms of operating scale, the lidar camp is ahead. Still, the fact that the sensor cost per vehicle differs between Waymo and Tesla—Karpathy's framing of "hardware problem versus software problem"—remains just as relevant. Waymo has to absorb the per-vehicle cost as it scales, while Tesla only needs to finish the software on cars already sold—though when that finish line arrives remains unknown.
So why is Matic still betting on cameras
In an era when lidar costs $200, removing it is no longer a cost-cutting strategy. And cost isn't the reason Matic gives, either.
The company's two co-founders come from Google Nest. CEO Navneet Dalal was lead author, during his time at INRIA in France, of a 2005 CVPR paper on "Histograms of Oriented Gradients for Human Detection"—a technique that served as the standard for pedestrian detection for nearly a decade before deep learning took over. Together with co-founder Mehul Nariyawala, he built facial recognition company Like.com (acquired by Google in 2010) and gesture recognition company Flutter (acquired by Google in 2013), and at Nest he shipped deep learning-based person and motion detection into products. It's a partnership that has spent 20 years tackling the problem of finding people with cameras.
The first reason they cite is, surprisingly, organizational. Based on a rule of thumb from their time at Nest, every additional sensor requires two to three dedicated software engineers to continually maintain it—calibration drifts, supply chains get disrupted, and failure points multiply. The second reason is more fundamental: homes are spaces people built for other people's eyes. The third is update speed: software can be swapped out, but hardware sensors cannot.
The cost was seven years. Founded in 2017, by fall 2020 the team judged that existing open-source SLAM solutions topped out at 70-80% accuracy and rewrote everything from scratch, then spent another 9-12 months switching processors from Qualcomm Snapdragon to NVIDIA. The company has raised $29.6 million total, with a notable list of individual investors including Nat Friedman (former GitHub CEO), Daniel Gross, and Stripe's Collison brothers.
Pricing has moved from a launch price of $1,495 down to $1,095, and currently sits at $1,245, though the company has announced it will raise the price back to $1,495 starting September 9, citing component tariffs. The next hardware model has been delayed. According to the company, rather than releasing new hardware, it chose to extract everything possible from the current hardware first—applying the same "buy back with software" strategy to its product roadmap.
Why this matters to Korean readers
Korea's robot vacuum market is approaching 1 trillion won annually, with Roborock holding the No. 1 spot for four consecutive years and around 50% market share. Its flagship models are priced around 1.9-2 million won. Samsung and LG's countermove hasn't been suction power but security—Samsung leading with Knox, LG with Shield. Now that devices using cameras and sensors to perceive indoor spaces have entered bedrooms, personal data protection has become a purchasing criterion.
This wariness is grounded in an actual incident. In December 2022, MIT Technology Review reported that photos captured by iRobot Roomba development test units had leaked, via an outsourced labeling company, into private social media groups—including a photo of a woman sitting on a toilet. This is the original sin that camera-equipped cleaning robots carry with them.
This is why Matic emphasizes on-device processing so heavily. The company has stated that "all processing happens on the device, and nothing leaves the four walls of the home." However, a crack has appeared in that principle with the voice feature unveiled on August 13. Voice commands beginning with "Hey Matic" are processed via Google Gemini's cloud servers and currently require an internet connection. The company has said its goal is local processing that delivers "80% of the benefit" without internet access. On-device processing, in other words, is more an aspiration than a finished state.
So the question left for Korean consumers isn't "are cameras cheaper than lidar?" Lidar has already become plenty cheap. The question narrows to two: what more can be done using cameras alone, and where does the 3D map of your home that those cameras capture actually end up?
So what actually changes
A lidar-free cleaning robot isn't, on its own, a novel technology. What makes Matic's case interesting isn't an experiment in cutting costs with cheap sensors, but a demonstration of what becomes possible once you first build a photorealistic map of the home—showing what can be built on top of it. A feature where the robot goes wherever you point is difficult to build from lidar point clouds alone; calculating direction requires seeing where a person's arm is pointing, and that's a camera's job.
At the same time, Karpathy's "march of nines" applies just as much to cleaning robots. Working well in a demo video and working well across thousands of other people's homes are different problems—which is exactly why Matic spent seven years on this. Now that the still-unresolved vision-only debate from autonomous driving has migrated into home robotics, what's worth watching isn't the fact that lidar was removed. It's whether the calculation of buying back what was removed with data and compute holds up in the living room as well as it does on the road.
