Ex-DeepRoute Executive Liu Nianqiu Launches Feimoyi to Build Open-World Physical Agents

Liu's route to embodied intelligence was unusual. He worked at Intel on new process and equipment research and development, then founded companies in IoT and industrial intelligence. According to Leiphone, he joined DeepRoute in 2018 and worked across hardware, Robotaxi operations, product, data, mass production and after-sales. In the Leiphone interview, Liu said he formally entered autonomous driving in 2019, after DeepRoute raised a large domestic financing round and invited him to join. The article and interview therefore give different years for his move to DeepRoute. From 2024, Liu followed embodied intelligence closely, reading papers and studying supply chains and technical routes. By the end of 2025, he judged that the timing for a startup had matured: robot supply chains were complete enough that a new company did not need to spend most of its energy on hardware and body weaknesses, and the AI-driven technology and data paradigm had become clearer. After deciding to start the company, he spoke with several DeepRoute co-founders, who supported another zero-to-one effort. The seed round came together quickly; when he first met investors, Liu had not prepared a business plan, the article said.

Feimoyi's main bet is on robots that can work in less-defined settings. Liu calls them Physical Agents, saying robots should eventually enter open living environments and complete tasks for different people at different times and places. The common embodied-AI path in recent years has been to choose clear-boundary scenes such as screwing in factories, delivery in hotels or moving goods in warehouses, then define the task and use engineering and models to raise success rates. Liu does not dismiss the commercial value of those scenes, but he questions how much capability a robot can transfer if it works for a long time in a highly defined environment. Feimoyi is therefore taking a route with blurrier boundaries: open, dynamic scenes. Commercial districts, office buildings, residences and event areas form the life circles it wants robots to enter gradually. A robot might fetch something, buy coffee, take out trash or help a courier when the courier is busy. It is not fixed into one occupation in advance; its task changes with the environment and demand. Feimoyi calls this role a partner. The more environments, people, objects and tasks a robot encounters, the more likely it is to form cross-scene transferability, Liu said. A key task this year is initial data collection, and the company plans multi-point deployment next year so that Physical Agents can form a loop of deployment, data, training and redeployment. The unresolved problem is that open-world complexity has almost no upper limit. A robot completing one of ten tasks in a demo and a robot completing tasks whenever a user calls it are different standards, he said.

Much of Feimoyi's approach draws on DeepRoute's mapless autonomous-driving experience. In the interview, Liu described how DeepRoute operated Robotaxi services in Shenzhen for more than a year, starting in downtown CBD areas. The team later decided that a high-definition-map-based approach had no future and moved to a mapless route. As a startup, it could not do both well or afford two lines, so it stopped Robotaxi and focused on mapless driving. Liu said he began thinking in 2021 about whether the business could work at scale; because he directly oversaw operations, he knew actual operating costs. He raised the idea of adjusting direction inside the team, and the core team discussed and decided the final direction together. The wind-down was gradual. After deciding to withdraw, DeepRoute reduced its fleet step by step rather than stopping all service on the day of the decision. It still had a small group of loyal users, and customer service communicated with them. Product work that had focused on Robotaxi operations shifted toward mass-production products and automakers. The team rethought what functions an L2 assisted-driving product should have, how it should operate and exit under different road conditions, which was a different set of questions from Robotaxi operations. Technically, moving from mapped to mapless driving amounted to a rebuild. The older mapped version was relatively traditional, with perception, prediction, tracking, high-definition-map localization and planning control. The mapless version greatly expanded perception boundaries, essentially absorbed map localization, and became tightly coupled with planning. Data requirements also changed, making it a large architectural shift.

Liu's responsibilities included adjusting the sensor scheme, since the mapless system relies more on vision, and building a data production system. In mass-production projects, the company had to reconsider how to choose and configure sensors to meet algorithm needs and mass-production requirements. It also had to decide how data would be produced, how to work with external partners and how to control costs. Liu said research projects could not set a rigorous deadline because no one could be sure when a result would emerge. The team focused on moving fast, validating paths and quickly adjusting. They defined model structures, decided how data would be collected, labeled and trained, then ran tests to see whether problems came from road understanding, planning, model capability or front-end data. If compute was available, they wanted it running while people rested. Core departments had few fixed rest days, with engineers rotating and working back-to-back, though final judgment tended to converge on a small number of people with clear understanding of the problems. It took about a year from the decision to go mapless to winning a production nomination, Liu said. Pressure increased after the nomination, because the next step was not proving the technology but mass-producing it. External doubt was reasonable at the time, Liu said, because traditional mapped systems performed better in areas already covered by high-definition maps. Many peers were themselves in the high-definition-map industry and naturally resisted the change. Looking back, Liu considers mapless driving an industry turning point. City assisted driving had once been usable in only a few cities; after mapless systems landed at scale, consumers could feel that the function worked where they themselves drove.

Liu also recalled a vehicle retrofit that preceded a production nomination from a leading technology company. The client had experienced DeepRoute's car and was interested in its capability and solution, but wanted to see how it would perform on a mainstream well-known model. The client provided a car without assisted driving and asked for a quick demonstration. The retrofit took less than two days. The car arrived without the relevant system, and the team had to drill holes, replace glass, add controllers and sensors, restore the vehicle, connect the algorithm and conduct road tests and evaluation. People worked in shifts for several days. Liu, responsible for hardware and delivery, had to turn the client's experience requirements into an actual vehicle by installing sensors and controllers, connecting the system and supporting testing and problem solving. The client saw a traditional car converted into an autonomous-driving test prototype in a short time, which demonstrated the team's execution and delivery ability and increased confidence in mass production. It was not mass production, but it let the client judge whether the team could face real vehicles and real problems. For Liu, the hardest part of moving from a mapless system that could run to one that could be mass-produced was system integration and coordination with automakers. In research and development, DeepRoute used a system it built itself; if the software ran in the R&D environment, testing could proceed. In mass production, the company had to define many engineering details with the carmaker, including diagnostic signals, signal matrices, middleware communication reliability, functional-safety exceptions and handling mechanisms, quality assurance, version management and after-sales interfaces. Automaker after-sales teams were more used to handling a broken component, while diagnosing assisted-driving system problems required both sides to build an after-sales diagnostic system through joint work. The process took the company through a full mass-production engineering cycle. Beyond algorithms, there were many system, quality and collaboration problems. The pressure, Liu said, was in opening these links with automakers so that developed capabilities could enter the mass-production delivery system.

Liu's involvement with autonomous driving began earlier than many might expect. He started working on assisted-driving hardware in 2017, helping Zhou Guang, DeepRoute's founder and CEO, and others. At that time, they focused more on algorithms and needed hardware engineering support. Liu had been running startups in industrial intelligence, and much of the early roof-mounted lidar, controllers and other equipment for their vehicles was made by his team. In 2019, after DeepRoute raised a relatively large single financing round in China, the company invited him to join. He thought he could learn deeply and formally entered the assisted-driving industry. Before that, his experience looked distant from autonomous driving. He had founded companies in IoT, industrial intelligence and automation, and at Intel he was responsible for global new process and new equipment development. The technologies were similar in essence: production-line inspection of materials and environments requires sensor systems and control systems, which share much with intelligent driving. When Liu joined DeepRoute, the company had a few dozen people, mostly key technical staff who had worked together before. He did whatever was missing. The core team mainly worked on perception, planning and map localization, while he was responsible for hardware and engineering and built the early hardware team. His first task was a hardware synchronization controller for different sensors on the vehicle. The car used mechanical rotating lidar, integrated navigation and industrial cameras, not today's automotive cameras. These devices came from different industries and had different operating modes: lidar had a scanning cycle, integrated navigation had its own frequency and timing requirements, and cameras had exposure and trigger conditions. To use them in one autonomous-driving system, they had to be triggered at specified times and have their data timestamps on the same timeline. The required synchronization accuracy was better than one millisecond. Liu personally studied the devices' characteristics, designed a circuit system, interfaced with the sensors and completed interaction and synchronization control, unifying multiple devices' timelines to the millisecond level. At that time, autonomous-driving R&D was divided into modules such as perception, planning, map and localization, and he was mainly responsible for hardware.