From Map-Free Autonomous Driving to Robotics: The Embodied AI Startup Journey of Feimo-One's Liu Nianqiu

Liu Nianqiu, a veteran of DeepRoute.ai, is steering his new company Feimo-One into the embodied intelligence space. His founding thesis is straightforward: the technical paradigms and engineering know-how accumulated through autonomous driving's shift from HD maps to a "map-free" approach can be transplanted directly into robotics. Right out of the gate, Feimo-One closed a nearly $100 million seed round led by IDG Capital, with participation from Yunqi Capital, DiDi, Fosun RZ Capital, and Hesai Technology — and when Liu first met with investors, he hadn't even had time to prepare a pitch deck.

An Atypical Embodied AI Founder

Liu Nianqiu doesn't fit the standard mold of an embodied AI founder. He's not an academic out of a robotics lab, nor a serial entrepreneur who only discovered Physical AI in the past two years. He previously led global R&D for new processes and equipment at Intel, launched his first startup in IoT and industrial intelligence, and in 2017 began helping DeepRoute.ai founder Zhou Guang with autonomous driving hardware. He joined DeepRoute officially in 2018, working his way from hardware all the way through Robotaxi operations, product, data, mass production, and after-sales service.

This "plug any gap" career trajectory is exactly what gives him confidence in tackling embodied intelligence. What he brought out of DeepRoute wasn't former colleagues — almost no one left with him — but rather the know-how: large-scale data flywheels, productionization, technical judgment, and roadmapping. "Having that project experience and then building a new embodied team — the logic is completely different."

Shared Technical DNA

Liu acknowledges that his earlier experience and autonomous driving "look pretty far apart on the surface, but are technically quite similar at the core." Doing assembly-line inspection at Intel required sensor systems paired with control systems — a foundation with a lot in common with intelligent driving. His first assignment at DeepRoute was building a hardware sync controller for mechanical rotating LiDARs, integrated navigation systems, and industrial cameras: getting devices from completely different industries, each with its own timing requirements, onto a unified timeline accurate to better than one millisecond.

DeepRoute's Pivotal Shift: Killing Robotaxi, Betting on Map-Free

Feimo-One's technical convictions are largely shaped by what Liu describes as DeepRoute's "tear it all down and rebuild" transformation. Around 2021, Liu — who was directly overseeing operations at the time — began questioning the real-world costs and scalability of Robotaxi. The core team reached a collective decision: the HD-map-based approach had no future, and all-in on map-free was the only path.

"As a startup, you can't execute two tracks well simultaneously, and you don't have the resources for it." DeepRoute gradually scaled back its Shenzhen fleet and shifted product focus from Robotaxi operations to L2 mass production and OEM partnerships.

What the Map-Free Shift Actually Means

This was an architectural overhaul. As Liu explains it, the old HD-map stack consisted of traditional perception, prediction, and tracking modules, combined with HD-map-based localization, feeding into planning and control. Going map-free meant dramatically expanding the scope of perception — absorbing map-based localization into the perception layer, tightly coupling it with planning, and completely changing data requirements.

During R&D, there were no strict deadlines in the traditional sense. "It's hard to set a rigorous deadline on a research project, because nobody can guarantee results by a specific date." The team's method was to define the model architecture, decide how to collect, label, and train on data, get it running, and immediately decompose any problems that surfaced: Was it a road-understanding failure, or did it understand correctly but plan incorrectly? "When people rest, compute shouldn't."

From the decision to go map-free to securing a design win took roughly one year. But Liu is quick to point out that winning the design award was actually when the pressure intensified — because the next step wasn't proving the technology could work; it was mass production. He found system integration with OEMs to be personally the hardest part: diagnostic signals, signal matrices, middleware communication reliability, functional safety exception handling, and building an entirely new intelligent driving after-sales diagnostic framework from scratch. "Beyond the algorithms, there are enormous systems, quality, and collaboration challenges to solve."

HD Map vs. Map-Free — HD Map solutions rely on pre-collected and continuously updated centimeter-precision maps; vehicles match sensor data against the map for precise localization, while the planning module computes paths on a known static road structure. This architecture performs reliably on closed or semi-closed routes, but maintenance costs are extremely high — every road change (construction, new lane markings, temporary closures) requires manual map updates, and geographic coverage directly caps the operational area. Map-free approaches require the perception module to understand road topology, obstacles, and drivable regions in real time, without relying on stored prior maps. This demands far stronger generalization from perception networks — handling road geometries never seen in training data — making the diversity and scale of training data the core competitive factor. The shift is essentially converting "the map's memory" into "the model's understanding" — and it's precisely the paradigm Liu believes transfers directly to robotics, since robots equally cannot rely on pre-built environment maps to operate in the open world.

Why Start a Company Now

Liu began seriously tracking Physical Agents in 2024, a period when Figure and Unitree H1 were making frequent appearances but mostly demonstrating walking and stair-climbing. By late 2025, he concluded the timing was right.

His reasoning had two pillars. First, the supply chain had matured — startups no longer needed to be dragged down by weak hardware and spend enormous energy playing catch-up. Second, the AI-driven technical and data paradigms were becoming clear enough to start building genuinely long-term-valuable foundations.

He has a precise sense of where the window is: before the end of 2025, the industry was mostly exploring task-specific solutions; after that, the conversation increasingly shifted to heterogeneous data, diverse data, and generalization. "Go in too early, and your resources get burned on specialization and exploration without forming a core foundation. Go in too late, and others have already built that foundation."

As for why he didn't stay at DeepRoute and build this from within, his answer is candid: embodied AI demands extreme resource investment and focus, and at DeepRoute he had already walked the full arc from 0-to-1, 1-to-10, and 10-to-100. "The things I'm good at were no longer generating much differentiated value at that stage. Rather than coast, I'd rather go do another 0-to-1 thing."

Physical Agent: A Partner, Not a Tool

Feimo-One deliberately uses the term "Physical Agent" rather than the more fashionable "Physical AI." Liu's view is that Physical AI describes a capability, not a productive force — whereas what they're building is an Agent in the physical world. Like a digital-world agent, it assists humans in completing tasks: with cognitive-space decision-making ability, action-space planning ability, contact-space detail-alignment ability, and memory. "The word 'Agent' is more direct for users. It's something that can help me get things done — not a concept."

Why Open, Dynamic Environments

The most common embodied AI deployment path over the past two years has been to find a clearly bounded scene first — tightening screws in a factory, delivering items at a hotel, moving goods in a warehouse — and drive success rates up from there. Liu doesn't dispute the commercial value of those scenarios, but his question is: if a robot works exclusively in highly defined environments for years, how much of what it learns actually transfers?

"If you're tightening screws in a factory every day, no matter how good you get at it, that doesn't mean you can go out on the street and deliver food tomorrow. There's no natural transfer capability built in."

Feimo-One has therefore chosen open, dynamic environments with fuzzy boundaries: commercial districts, office buildings, residential areas, event spaces — the "living circles" robots will gradually enter. It delivers food today, packages tomorrow, handles cleaning the day after — it doesn't get locked into a single occupation. The richer the environments a robot enters, the more likely it is to develop cross-scenario transfer capabilities.

The "Partner" Relationship and Hardware Choices

The reason for calling it a "partner" rather than a tool is that early on, humans and robots help each other: a robot that can't find its way a kilometer out can have the user guide it through its eyes; a robot that can't tell which bag at the restaurant is yours can have you identify it. As capabilities improve, the proportion of human-assists-robot decreases while the proportion of robot-assists-human increases.

The product retains strong humanoid characteristics in service of this same logic: to do all kinds of tasks and get itself to where those tasks need doing, it needs two hands, a head, visual height close to a human's, and legs capable of navigating everyday stairs and minor obstacles. "The core purpose isn't to look like a person — it's to enable it to enter these environments and complete tasks the way a person would."

Rejecting Engineering Shortcuts, Staying Data-Driven

Feimo-One has no plans to rely on pre-built maps or large-scale pre-deployment staging, as early autonomous driving companies did. "We already moved away from HD maps in autonomous driving — Physical Agents definitely can't use them either." Liu sees this as central to being AI Native: early autonomous driving wrote enormous volumes of rules, which worked when the operating domain was small, but as the domain expanded, adding each new rule could break other scenarios. With tens of thousands of users you'd still be maintaining the same rule set. "It eventually transforms from an asset into a liability."

Agent in AI refers specifically to a system that can perceive its environment, autonomously plan, and execute multi-step tasks — distinct from tool-type models that only execute single instructions. Digital-world Agents (such as LLM-based automation assistants) can already handle composite tasks like web browsing, code writing, and information synthesis; their core capability lies in decomposing high-level goals into executable steps and dynamically adjusting based on feedback during execution. Physical Agents extend this logic into the physical world: rather than receiving precise instructions like "move the box from A to B," a robot understands fuzzy goals like "go get my delivery," autonomously planning the complete sequence — go outside, identify the package, carry it back — and making autonomous decisions when it encounters obstacles (elevator out of service, wrong locker code). This requires the system to simultaneously handle semantic understanding, spatial reasoning, and fine motor operation, and coordinating among all three is one of the central technical challenges in embodied intelligence today.

Closing the Loop and the Industry's Unsolved Problems

On the current roadmap, Feimo-One will release its own hardware platform this year, layer on a large model, run initial deployment demonstrations, and begin bootstrapping data collection using self-developed collection equipment. Starting next year, the plan calls for multi-site deployment and genuinely closing the "deploy → data → train → redeploy" flywheel.

Liu repeatedly comes back to a lesson from autonomous driving: the chasm between demo, productionization, and real deployment. "Getting it right once out of ten tries, or picking the one best-looking scenario out of ten to film, is completely different from a user being able to show up at any time and have it work." The complexity of the open world has essentially no upper bound — this is a problem the entire industry hasn't solved yet. The focus isn't on the ceiling of single-point capabilities; it's on the floor of performance across a range of scenarios.

Asked about his vision three years out, he didn't cite market share or headline numbers. He simply said he hopes that in many places, people are genuinely using their "partners." "At its core, it's still about whether the product earns real recognition." That statement — from the first user who actually hailed a Robotaxi, to the mass-production system handed off to car owners, to robots today — has been his consistent benchmark throughout.

Data Flywheel is an engineering paradigm validated by the autonomous driving industry: long-tail scenarios encountered during real-world deployment are captured and transmitted back, filtered and labeled into the training set, iterated into model updates pushed back to vehicles, forming a virtuous cycle of "broader deployment → more data → stronger models → broader deployment." This paradigm is instructive for embodied intelligence because the complexity of the physical world equally cannot be covered by simulation or limited lab data. However, the data flywheel faces a higher bar for robots than for autonomous driving: autonomous vehicles can naturally generate massive driving mileage through production-scale fleets, whereas robot deployments start small and each robot covers scenario types far more fragmented than a car driving on roads. This makes designing efficient data collection and filtering strategies for the early small-scale deployment phase a core engineering challenge embodied intelligence companies must solve proactively.