Over the past two years, global attention has fixated on generative artificial intelligence and humanoid robots. The former is reshaping how humans interact with information, while the latter is viewed as a potential pathway toward general intelligence.
Yet beyond the hype, the robots already delivering measurable productivity gains have been quietly working on city streets. They can be found by the curbs, beneath overpasses, and in the corners of sidewalks that most people tend to overlook. Their shifts begin around five in the morning, taking on tough, dirty, and often risky tasks once done by humans.
In China, sanitation is a massive industry that continues to face longstanding challenges. It is labor-intensive, difficult to staff, and fraught with safety risks. Standards are high, but the work is anything but glamorous. Precisely for this reason, sanitation offers one of the clearest ways to measure the real-world value of AI.
That is why CowaRobot chose sanitation as its starting point. “Our approach is pragmatic,” said Liao Wenlong, the company’s CTO. “We center our work on urban services, starting with sanitation. By tackling the dirty work of cities, our robots free up human labor.”
In October, CowaRobot introduced the R0, a compact robot designed for municipal operations, with plans to adapt it later for more complex environments such as property management. Around the same time, the company won first place at the Shenzhen International Intelligent Sanitation Cleaning Robot Competition, validating its innovation and commercial readiness.
Finding a landing pad
To most people, the logic seems simple: if a household vacuum robot can clean a living room, a larger one should be able to clean a street.
In reality, that assumption breaks down quickly. Robots that interact with the physical world require two core capabilities: navigation, to move autonomously, and operation, to perform specific tasks. “From that perspective, robotaxis, sanitation vehicles, and home cleaning robots are simply different combinations of those two abilities,” Liao said.
Once placed in real environments, however, the differences become striking. These extend from perception and decision-making to task execution and safety parameters.
Household cleaning is a matter of convenience. Street cleaning, by contrast, is essential infrastructure. Labor shortages and high attrition rates plague sanitation departments, yet cities cannot compromise on cleanliness.
This was CowaRobot’s starting point. “We want embodied intelligence to genuinely improve productivity,” Liao said. “Our sanitation robots are designed to be productive now while advancing toward the ultimate goal of generalized physical AI.”
Among current applications of embodied intelligence, sanitation offers both the most viable commercial entry point and one of the steepest technical barriers. It is a typical B2B scenario: high frequency, non-optional, and easily measurable. Yet it also exposes four of the toughest challenges for any physical AI system:
- Unstructured environments: Curbs, sidewalks, and greenbelt edges lack clearly defined drivable zones. Humans navigate them by instinct, while robots must learn to interpret spatial relationships and intent directly, pushing development from modular systems toward end-to-end world models.
- Dynamic safety decisions: Navigating intersections, avoiding pedestrians, and negotiating with cyclists require predictive reasoning about future outcomes, not just reactions to current positions.
- Precision under tight constraints: When sweeping near curbs, even a few centimeters matter. Too far and dust remains; too close and the machine risks scraping a pole or the curb itself.
- Control coupling: Sanitation vehicles must drive, brush, and track moving debris simultaneously. Every minor adjustment affects the result.
Few companies can master all four dimensions. Those that oversimplify fail to replace human labor, while those that pursue full autonomy face years of research and high capital costs.
When AI enters the physical world
Unlike autonomous vehicles, which mainly travel between two points, sanitation robots must move while working. They must understand space, manage tasks, and adapt to constantly changing conditions simultaneously.
For years, most developers have relied on decoupled frameworks: perception detects obstacles, prediction handles trajectories, and control executes preset rules. In open, dynamic environments, however, such architectures are brittle and hard to scale.
CowaRobot is rethinking that approach. According to Liao, industry capabilities have evolved through five stages:
- Scripted autonomy, in which robots follow fixed routes within closed environments.
- Map-based navigation, which guides robots using high-precision maps to operate along public routes.
- Map-light adaptation, enabling real-time adjustments without relying strictly on high-definition maps.
- Physically deployable AI agents, which operate autonomously out of the box, planning routes and tasks without degradation.
- Cloud and edge integration, enabling coordination among multiple embodied robots for optimized resource allocation.
Liao said CowaRobot is currently at the fourth stage and is progressing steadily toward the fifth.
Its system is built around a bird’s-eye view (BEV) world model that predicts uncertain future states and generates actions directly, mimicking human intuition. “When the wind blows trash, the robot anticipates where it will drift,” Liao said. “When it’s too close to a wall, it knows the consequence of collision.”
To handle complex urban features such as traffic lights, no-parking zones, and tactile paving, CowaRobot integrates a visual language model (VLM) as a cognitive layer alongside the physical model. “Think of the VLM as the robot’s brain,” Liao said. “It reasons deeply when needed, then guides motion.”
Through reinforcement learning, this architecture enables training in simulated environments, covering both rare edge cases and coordinated multi-action behaviors.
To make its systems plug-and-play, CowaRobot added two key features. One is a self-memory mechanism that allows robots to recall and adapt to environments over time, improving performance with each pass. The other is prompt-based behavior adjustment, which enables adaptation to local conditions without retraining.
A decade of data, hardware, and insight
In embodied intelligence, where deep engineering and long-term commitment are crucial, it is said that time builds moats rather than pressure.
He Tao, CowaRobot’s founder and CEO, holds degrees from Shanghai Jiao Tong University and the Tokyo Institute of Technology, where he also taught. Liao’s background spans control theory and AI, and much of the team shares similar technical depth.
A decade ago, when many companies were racing to deploy robotaxis, CowaRobot took a different route, focusing on perfecting robots in real urban settings first. Over ten years, it has built advantages across three dimensions:
- Hardware control: From design to manufacturing, CowaRobot develops its own systems, similar to Unitree Robotics’ approach to motor and motion control.
- Data scale: The company has reportedly collected 50 petabytes of real-world driving data across sidewalks, parks, and auxiliary roads, which are areas rarely covered in passenger vehicle datasets. It also employs self-supervised learning and VLM-assisted automatic labeling.
- Operational insight: A decade of field experience has shaped its understanding of task logic, including when to sweep or spray, when to break minor traffic rules for efficiency, and how to identify illegal roadside waste.
A broader market and a bigger opportunity
The clearest test of technological maturity is whether the commercial loop closes. Can the system replace human labor, scale across cities, and maintain efficiency over time?
By investing heavily in R&D, CowaRobot believes it has achieved a balance between cost, performance, and adaptability.
Under ideal conditions, one sanitation robot can cover 20–30 kilometers per day, equivalent to the workload of five to ten human cleaners. Even at a modest labor benchmark of RMB 30,000 (USD 4,200) per worker annually, each robot yields a positive gross margin.
Hardware costs have fallen by more than 70 percent from the first-generation model, aided by vertical integration from chassis to cleaning modules. The company also replaced expensive LiDAR (light detection and ranging) sensors with vision-based systems, improving cost-efficiency while maintaining autonomy.
In short, CowaRobot’s advantage lies not only in cost-effectiveness but in AI-driven productivity.
Today, its robots operate routinely across China and are expanding into Singapore and the Middle East. Future applications include litter collection in greenbelts, graffiti removal, and trash can maintenance.
Its new R0 hybrid robot, equipped with dual arms, is already undergoing object pickup tests and is expected to enter the property services and consumer markets soon.
Sanitation automation is far from a niche market. China’s Ministry of Finance reported that in 2024, urban and rural environmental sanitation expenditures reached RMB 242.6 billion (USD 34 billion), with continued growth expected.
As urbanization accelerates, populations age, and labor costs rise, cities will increasingly rely on scalable, intelligent systems to maintain hygiene standards.
With robust engineering talent, complex urban environments suited for training, and vast domestic demand, CowaRobot is well-positioned to help set global benchmarks for embodied intelligence.
KrASIA Connection features translated and adapted content that was originally published by 36Kr. This article was written by Xiao Xi for 36Kr.