DEEP Robotics Embodied AI Research Lands Amid $367M IPO Bid

Five research projects, one stated goal

On October 24, Hangzhou-based DEEP Robotics announced five research projects. The company says they all feed DEEPRobotics-Brain, the architecture it is building to control robots doing real work in the field.14 The five are a world model (DR-WM-1.0), an embodied foundation model (DR-VLA-2.0), a reinforcement-learning framework (DR-APO), an embodied agent with memory that updates as conditions change (DR-Agent), and a platform for data and evaluation (DR-Benchmark).11 The company describes a single goal behind all of them: getting robots to keep working when their surroundings change.11

The list of collaborators is long. It includes MIT, UC San Diego, the University of Illinois Urbana-Champaign, Tsinghua, Zhejiang University, the Hong Kong University of Science and Technology and the Beijing Institute of Technology.11 The preprints are posted on the company's developer portal.12

The release reads like an academic summary, but the timing matters. Five months earlier, DEEP Robotics applied to list on Shanghai's STAR Market, and the largest share of the planned proceeds is for building robot "brains."68 This research announcement is best understood as a public preview of that spending.

What each project claims

DR-WM-1.0 deals with a known weakness of video-generating world models: a predicted clip can look realistic and still be physically impossible. The company's example is a convincing video of a robot crossing snow that never shows the slipping low friction would cause.11 Its approach combines reconstructed scene geometry with physical estimates of friction, roughness and stiffness, calibrated against real data on how the robot responded to its actions.12 Training used 2.58 million image-mask pairs and more than 100,000 clips from real robots.11 The suggested use is substation inspection after rain or snow, where a robot could simulate a slippery shortcut and a longer, safer route before picking one.11

DR-VLA-2.0 covers both navigation and manipulation. For navigation, the team built EgoLynx, which pairs 10,000 hours of curated first-person web video with 2,000 robot navigation clips collected by teleoperation.11 The bet is that human footage can teach a robot how to understand and predict a scene, while the smaller robot dataset teaches it which actions it can actually execute.11 The reported gains are modest: about 5% on a combined navigation score and a 14.6% cut in final heading error, each measured against the best competing result.11

The manipulation part, called Atomic Motion Coordinate (AMC), breaks each arm's movement into 13 interpretable "motion atoms." Training first aligns language with geometric motion before visual input is added, because in many existing vision-language-action (VLA) models the visual signal can drown out the instruction.11 AMC reached 87.8% normalized task progress on bimanual tasks with unseen layouts. With force feedback, it reached 78.5% on a drawer-and-plug task and 75.2% on vase wiping. Baselines included a fine-tuned π0.5.11

DR-APO was built with Zhejiang University. It separates the policy-improvement step from the way a policy is represented, so the same method works for Gaussian, diffusion and flow-matching policies.13 The most notable result comes from real hardware. Applied to an existing navigation policy on the company's LYNX M20 wheel-legged robot, it raised real-world success from 25% to 65%.11 In simulation, success rose from 70.33% to 89.55%.11

DR-Agent addresses a practical problem: a robot's memory goes stale when objects are moved while it isn't looking.12 Its EvolvingNav method keeps a time-stamped estimate of where things are likely to be and updates it with a predict-observe-replan loop. It is evaluated on a new benchmark of 54 simulated scenes and 803,680 task instances.11

DR-Benchmark is built on LynxScenes, a dataset recorded with the LYNX M20 over 833.6 hours and 3,294.7 kilometers of outdoor operation. It contains about 100,000 navigation clips and 3 million annotated frames.11 The benchmark runs 25 models on four tasks that share the same data: trajectory prediction, bird's-eye-view perception, future scene prediction and scene reconstruction.11

Reading the results critically

The announcement shows a company that has studied the main open problems in robot foundation models: physically grounded world models, moving human video into robot policies, and VLA models that actually follow language. Academic reviews of embodied AI name similar priorities, including training aimed at understanding the physical world, predicting the outcomes of actions, and memory that lasts over long tasks.1618

The limits deserve equal attention. Every number above comes from DEEP Robotics' own announcement, and the work is released as preprints rather than peer-reviewed papers.12 Several of the reported gains are measured against baselines the authors chose, or on benchmarks the company built. That doesn't make the gains less real, but it does mean outside researchers should reproduce them before treating them as settled. The most convincing items are the hard-to-fake real-robot results, especially DR-APO's jump from 25% to 65% on hardware. The weakest are the simulation scores.

Humanoids are mostly absent from the announcement. Its data, tests and example uses center on wheel-legged and quadruped robots and on industrial settings such as substations, electrical cabinets and delivery routes.11 The manipulation work does involve two-armed tasks, and the company says the research program is meant to work across different robot bodies.11 For now, though, these results apply mainly to legged robots in the field, not to general-purpose humanoids.

The money behind the models

The financial background explains why DEEP Robotics is publicizing its research so heavily. Its STAR Market application was accepted in May. The company aims to raise 2.5 billion yuan, which coverage converts to between about $345 million and $367 million depending on the exchange rate used.56 One analysis puts the company's valuation above 10 billion yuan, or roughly $1.5 billion.3 About 1.169 billion yuan of the proceeds, close to half, is set aside for research on embodied algorithms and models.8 Read that way, the five projects show investors what the R&D portion of the IPO is supposed to pay for.

Coverage agrees on the main figures:

• Revenue: 337 million yuan in 2025, up from 103 million in 20246
• Profit: first annual profit of 28.7 million yuan3
• Gross margin: up from about 33.5% to 52.8%3

Outlets read those figures differently. Humanoids Daily presents the company's results as evidence that industrial robots aimed at hazardous jobs can be profitable.3 Caixin is more cautious. It reports that government subsidies of 11.9 million yuan made up more than a third of the profit, that inventory rose to 124 million yuan, and that the company sold only four humanoid robots in 2024 and 2025.6

HelloChinaTech makes a comparison that sharpens the point. Crosstown rival Unitree reported 1.71 billion yuan in 2025 revenue, about five times DEEP Robotics' figure.5 The same analysis argues that DEEP Robotics shows where robots really sell: environments where rough terrain, safety and reliability matter more than general intelligence. In its view, general embodied AI is still a bet that depends on raising capital.5

The sources also disagree on private funding details. One reports the December 2025 Series C as $71 million1, another as $68 million3, and Caixin describes it as more than 500 million yuan, co-led by CMB International and China Asset Management.4 A database tracker lists $140 million raised in total across four rounds.2 The gaps probably come from currency conversion and incomplete records. None of them changes the larger picture: DEEP Robotics is a well-funded company with state-linked investors, and it is now trying to raise money from public markets.

Why it matters

The research and the IPO filing make the same argument from two directions. The prospectus says most revenue still comes from quadrupeds doing inspections; industry applications made up 79.33% of main business revenue in 2025.5 The research builds on that business. Its largest real-world dataset was recorded on the company's own commercial robot, and its examples are taken from the substations and industrial sites where its customers already deploy robots.115 This is the company's clearest advantage. Its existing fleet produces real-world data that labs working only on humanoids find hard to match.

Our reading is that DEEP Robotics' most credible path in robot foundation models is to improve the robots it already sells, not to jump straight to humanoids. The five projects are serious and well connected, and the real-robot gains are significant. Whether the company can sustain margins above 50% while moving into crowded humanoid competition against Unitree, AgiBot and Western companies such as Figure is still an open question.3 This announcement doesn't settle it. What it does show is that the company's R&D spending is aimed at specific, measurable problems.