Do Vision-Language Models Really Help Cleaning Robots?

Short answer: what do vision-language models add?

Vision-language models, or VLMs, give a cleaning robot a layer of scene understanding. A LiDAR-based robot measures distance and geometry very well, but it does not know what an object is. VLMs aim to close that gap.

The honest answer is that the benefit is plausible and partly shown in research, but not yet proven on a working commercial floor. Vendors make strong claims about uptime and reduced human intervention. The independent evidence in our research comes mostly from academic navigation studies on legged robots.

For a facility manager, the useful questions are narrow. Which failures does the extra sensing fix, which new failures does it create, and what does it add to the cost of the machine and its upkeep?

What the March 2026 SoftBank launch actually claims

On March 4, 2026, SoftBank Robotics America announced AI-enabled commercial cleaning robots, according to its own press release. The lineup is Omnie for large-area scrubbing, V40 2.0 for vacuuming medium-to-large indoor spaces, and Phantas 1.3 for sweeping and light scrubbing in smaller areas.

The company says the robots use computer vision, vision-language models and 3D LiDAR to improve navigation, localization and obstacle avoidance. It also cites what it calls superior spatial awareness, which it links to higher uptime and less need for human intervention. Those are vendor-reported claims, and our research found no independent uptime figure to check them against.

Trade coverage names Gausium as the manufacturer behind the lineup, and SoftBank names senior living, hospitality, aviation and retail as target sectors. An aggregator article says Gausium Omnie units are operating at Milan airports for the 2026 Winter Olympics. That report was not checked against a primary source, so treat it as unconfirmed.

Why LiDAR alone leaves a gap

LiDAR builds a precise map of walls, furniture and open floor by timing laser returns. It is strong on geometry. One consumer review puts the limit plainly: LiDAR sees geometry, not identity. A chair leg and a cardboard box can look like the same obstacle shape.

Cameras add identity but bring weaknesses of their own. Consumer buyer guides note that cameras struggle with depth perception in some conditions, and another source warns that camera systems may misjudge distance or read mirrors and glass poorly. Those are consumer-product claims, so treat them as directional context rather than lab results.

That tradeoff explains why many flagship consumer vacuums now combine LiDAR with a camera. Commercial machines are following a similar logic, which is why the sensing mix deserves close questioning.

What the research shows VLMs can add

The strongest published result in our notes is a vision-language navigation study on a real quadruped robot. Single-target success rose from 66.7 percent for the vanilla system to 91.7 percent with the authors' intervention. Multi-target success rose from 41.7 to 75.0 percent, and unsupported target proposals per episode fell from 0.92 to 0.25 in the single-target case.

A quadruped is not a floor scrubber, so those numbers do not carry over directly. They do show the mechanism: language and vision reasoning can reduce wrong choices. A separate legged-robot study found that vision-language models can recognize scene states such as whether water is running or how clean a kitchen looks, which sits closer to the cleaning task.

Vision is not automatically the better obstacle sensor, though. A quadruped test found that a LiDAR-trained obstacle policy avoided obstacles the high-level vision-language-action model missed, while a blind policy got stuck. A multi-function floor robot prototype reported 100 percent obstacle avoidance in its own tests, with sensitivity to reflective surfaces that LiDAR helped mitigate. Those are authors' own tests, not field trials.

Where the evidence runs out

An indoor navigation benchmark for VLMs logged 659 path hallucinations, 418 obstacle hallucinations and 191 dimension-neglect errors across 1,500 question-and-answer pairs. An obstacle hallucination means the model reports a blocking object that is not there. Dimension neglect means it misses a narrow door or a tight clearance.

A phantom stop costs uptime, and a missed clearance risks a collision. That benchmark tested general indoor navigation, not cleaning robots, so it is a warning sign rather than a verdict. Still, it is the kind of failure a sales demo rarely shows.

Several things remain unproven for cleaning specifically. These include whether VLMs reliably recognize signage and floor hazards in real aisles, whether manual interventions fall on live routes, and whether uptime improves once the added sensing is maintained over months. SoftBank's uptime and intervention claims are its own, and we could not verify them.

Questions to put to any vendor about failure cases

Ask for specifics rather than adjectives. Vendors that have field data should be able to answer these without a slide deck.

• What does the robot do when the vision model and the LiDAR map disagree, and which one wins?
• What is the logged false-stop rate on a floor like ours, and can you share logs from a comparable site?
• How does performance change in low light, in glare from polished floors, and near reflective glass?
• Does the vision model run on the robot or in the cloud, and what happens during a network outage?
• What happens to route accuracy when a camera lens is dirty, scratched or misaligned, and who cleans it?
• How often are models updated, and does an update require re-mapping or re-validation of the site?
• Which failure modes trigger remote triage, and how quickly can on-site dispatch reach a stalled unit?

What the extra sensing costs beyond the sticker

Added sensing adds cost in three places: the hardware, the compute needed to run vision models, and the upkeep of cameras and software. Counterpoint Research says the commercial cleaning market achieved high-double-digit growth in 2025. That signals demand, but it says nothing about the economics of a single machine.

Price the whole picture: software terms, camera service, the robot maintenance service plan, and the downtime each failure mode causes. Ask whether a vision-equipped model and its LiDAR-only sibling carry the same service coverage.

Service Robot Co. leases robots through financing on a three-year lease, with one fixed payment quoted after a site assessment. As a vendor-neutral robot integrator, we compare units from several manufacturers against your floor, then handle robot deployment and integration, training and service through a nationwide US engineer network, so one partner covers the full lifecycle.

How to test the claim before committing

Treat the vision claim as a hypothesis for your floor. Run a timed robot pilot program on a real route, log every stop and manual intervention, and compare false stops and recovery times rather than headline percentages. If a vendor offers a LiDAR-only baseline, include it.

Site assessment mapping is the right place to set those measures. It is also where go-live support should begin, so the first weeks reveal the failures that a demo hides.

If the added sensing holds up in cluttered, mixed-traffic spaces, it may earn its place. If it does not, a well-mapped LiDAR robot with strong service coverage may serve you better.