Cainiao CTO Li Qiang founded a Physical AI platform startup, which has raised a seed round financing of over 100 million yuan from Yunqi Partners and SenseTime | Exclusive Release by Hard Krypton
Author | Huang Nan
Editor | Yuan Silai
Hard Krypton has learned that Physical AI platform company Quantum Dynamics recently completed a seed round financing of over 100 million RMB. This round was jointly invested by Yunqi Partners and SenseTime. The funds will be mainly used for core technology R&D of Physical AI, talent echelon construction and global market expansion, to accelerate the full-link construction of its physical world-oriented intelligent system from underlying models to scenario-based implementation. Multi-dimensional Capital participated in project incubation and team building.
Quantum Dynamics was founded in the first half of 2026. The company focuses on the deep integration of AI and the real physical world, self-develops general intelligent system technology that can be reused across multiple scenarios, and builds a Physical AI underlying platform for all industries.
At present, founders in the embodied intelligence track are mainly divided into several categories. One category comes from autonomous driving background, with strengths in system integration and engineering implementation; the other category comes from Neo Lab, with cutting-edge technical perspectives on embodied intelligence. Li Qiang, the founder of Quantum Dynamics, belongs to the third category. He has long been deeply engaged in logistics commercial scenarios, and has both AI technology understanding and global business operation experience.
Li Qiang, the founder and CEO of the company, has served in Alibaba Group for 17 years. He has undertaken the system construction, operation and management work from 0 to 1 in four sectors: Taobao & Tmall, Cainiao, international business and autonomous vehicles. He once served as CTO of Cainiao Group and CTO of Alibaba's international digital business sector, managing thousands of R&D teams distributed around the world; at the same time, he was concurrently the general manager of Cainiao Unmanned Vehicles, coordinating the R&D and commercialization of autonomous driving technology. His team has carried out in-depth business connection with postal systems in 200 overseas countries, accumulating first-hand experience in global logistics fulfillment.
Li Qiang, Founder and CEO of Quantum Dynamics (Source: Enterprise)
The core team covers all key links from cutting-edge research, model R&D, hardware engineering to global commercial implementation, and has the complete closed-loop capability from technological breakthrough to value delivery. This composite team gene also allows Quantum Dynamics to choose a differentiated route at the beginning of its establishment.
"We don't need to wait for a perfect general model to appear before looking for scenarios. Instead, we first find the entry point that can transform existing technologies into customer value, and take the lead in getting through commercialization," said Li Qiang, Founder and CEO of Quantum Dynamics.
In fact, since the beginning of this year, the embodied intelligence track is experiencing a collective shift from "showing off skills" to "pragmatism". At the just-concluded WAIC 2026 conference, the popularity of Demo booths that perform somersaults and dancing is still high, and the real-scene demonstrations that move automobile factories and logistics sorting lines to the site also attract crowds to stop and watch.
Industry consensus is forming that Physical AI has become the next main battlefield after large language models, and the winning factor of the competition is moving from the size of model parameters to the thickness of real data.
According to IDC report data, in 2025, China's user expenditure on embodied intelligent robots is expected to exceed 1.4 billion US dollars, and will surge to 77 billion US dollars by 2030, with an average annual compound growth rate of 94%. The faster the capital flows in, the more prominent a structural problem becomes: data in the physical world cannot be obtained in batches through crawlers like text and images.
This means that whoever first has the ability to deploy real machines on a large scale will have the priority of data collection. This is also the underlying logic for Quantum Dynamics to enter the industry: first go deep into the logistics warehouse, accumulate exclusive real data through real machine deployment, and then polish the underlying world-action model in reverse.
In a medium-sized warehouse, a picking robot completes tens of thousands of grasping, placing, scanning and labeling actions every day, covering almost all physical operation types from rigid boxes to flexible parcels, from standard sizes to special-shaped SKUs. Every success or failure of the action is a clear signal that does not require manual annotation. These are the links with the largest labor gap and the highest proportion of labor cost in the logistics network, and also the hard bones that traditional automation has been unable to tackle for a long time, which precisely requires model-driven solutions to achieve large-scale breakthroughs.
Targeting the logistics warehousing scenario, Quantum Dynamics builds a data closed-loop system (Source: Enterprise)
In terms of data collection and utilization efficiency, Quantum Dynamics has built a set of data closed-loop system. The end-side robot body runs lightweight algorithms, which can identify high-value data fragments in real time, including scenarios requiring manual intervention, "near-failure" states with low model prediction confidence, and difficult cases that have not succeeded after repeated attempts. These data will be automatically labeled, transmitted back, and enter the training pipeline.
At the same time, for the part that needs manual intervention in data processing, its algorithm platform will only intercept the most learning-valuable fragments in the intervention operation, eliminate redundant parts, to ensure that every frame of data entering the training set contributes effective information.
This mechanism makes Quantum Dynamics' data pipeline have two distinct characteristics. The first is the scale effect: every newly deployed warehouse adds a new data source; the second is the production of exclusive data. As the data becomes thicker, the generalization ability and success rate of the model become higher, and the cost and cycle of accessing new customers and new scenarios afterwards become lower.
When the real scenario data reaches sufficient density, it provides practical conditions for Quantum Dynamics' self-developed basic model. Its goal is not to build a single-point capability model, but to build an embodied intelligent system that can continuously evolve from real deployment. This system platform is composed of six technical modules working together, covering the complete Physical AI system of data aggregation, model update and cluster collaborative evolution, which will be verified and polished relying on offline real industrial scenarios.
The first is the embodied native World-Action Model. Different from common video prediction models, through joint modeling of vision, action, ontology state and tactile signals, the model can learn "how one operation will change the physical world", and meet the engineering requirements of real-time robot control in this process.
At the same time, based on the data closed-loop system that Quantum Dynamics has built, taking real machine operation and human operation data as the main raw materials, the original logs in logistics scenarios can be automatically processed into high-value training samples with task stages, failure causes, recovery actions and business results, supporting second-level retrieval, automatic cleaning and labeling, and day-level model iteration. Cooperating with the Human-in-the-Loop (HITL) mechanism, it can judge the timing of manual intervention of the robot, and convert each intervention into training data, so that the number of robots that a single person can manage continues to increase.
At the execution level, the system deploys a lightweight World Model Corrector, which is used to monitor action execution deviation in real time, and trigger deviation correction, re-planning and retry in time before the operation fails, so as to improve the task success rate in real scenarios. Facing new customers and new warehouses, the system does not need to retrain the full model weights. Instead, through Test-Time Training, few-shot demonstration and historical experience reuse, the robot can quickly adapt to different changes of goods, shelves, light and processes.
Finally, all embodied robots form a cluster that keeps learning and self-evolving. The success, failure, intervention and recovery experience generated by each device is continuously updated to the entire fleet through multiple processes of offline reinforcement learning, preference optimization, simulation training and real machine verification.
Quantum Dynamics divides its implementation rhythm into two stages. In 2026, it will focus on B2C and small and medium-sized B e-commerce warehouses, focus on the three labor-intensive processes of replenishment and shelving, goods picking, and parcel packing, prioritize the implementation of warehousing environments with regular SKUs of cosmetics and pharmaceuticals, and adopt the human-machine collaborative mixed production mode to complete large-scale pilots. By 2027, the business boundary will be further extended to complex express outlets, following the expansion path of "from logistics warehousing to industrial production lines, retail shelves, pharmaceutical sorting, to commercial services, and finally entering household people's livelihood services" to broaden the scenario boundary of Physical AI implementation.
The foothold of the Physical AI competition is never just the amazing degree of technical Demo, but whether the robot can stably complete operations in the real world. Quantum Dynamics has chosen a not-so-sexy path: first let thousands of robots run in the warehouse. It wants to prove not only the commercial feasibility of one company, but also the route feasibility of Physical AI "entering from vertical scenarios and gradually moving towards general purpose".
The following is an excerpt of the interview between Hard Krypton and Li Qiang, Founder and CEO of Quantum Dynamics (slightly edited):
Hard Krypton: Quantum Dynamics chooses to take logistics warehouses as natural training grounds, builds its own data flywheel through real machine deployment, and implements first before polishing the underlying model. Can you break down the time nodes and trade-off logic of this step-by-step strategy of "scenario first, then base"? What differentiated barriers can this path form in the long run?
Li Qiang: Our team has been deeply engaged in the logistics industry for many years. We not only clearly understand the current technical boundary of embodied intelligence, but also have a thorough understanding of the huge and complex real operation environment of logistics. Based on the current technical level, we have found a more pragmatic implementation path: relying on post-training solutions, we can first achieve large-scale commercial use in some warehousing scenarios.
The most direct benefit of doing so is that we can generate revenue while implementing, and continuously accumulate exclusive real machine data at the same time. These data are completely consistent with the distribution of the real physical world. They not only include the standard action trajectories of successful grasping, but also include a large number of negative samples such as grasping deviation and operation error. The richness and authenticity of the data are far higher than simulation and motion capture data sets. Under the same model architecture, relying on higher-quality real operation data, the generalization ability of the model will naturally widen the gap.
In the long run, the core of track competition will sooner or later fall on the efficiency of data collection and the cost of unit data. Whoever can continuously produce massive real machine data at lower cost can establish a solid advantage. This logic is consistent with that of Claude and Tesla FSD: the larger the scale of deployed terminals, the more accumulated interaction and driving data, the stronger the overall performance of the whole cluster after optimization through reinforcement learning and fleet iteration. Our large-scale deployment of warehousing robots is precisely to open up the complete closed-loop of real machine RL (Reinforcement Learning) post-training and continuous evolution of the whole fleet.
Hard Krypton: Quantum Dynamics plans to gradually expand from e-commerce warehouses to diversified scenarios, but the physical interaction logic and operation constraints in different environments vary greatly. What general atomic capabilities will the team precipitate from logistics data to support cross-scenario migration? For non-standard scenarios such as home and elderly care with weak commercial returns, what core difficulties remain to be broken through for the current model and few-shot adaptation framework?
Li Qiang: We have built three sets of layered evolution paths, all of which rely on the warehousing scenario to complete capability precipitation, and can be directly reused and expanded to various new scenarios later.
In terms of mobile operation, we follow a gradual polishing rhythm: first implement the desktop fixed manipulator solution, then realize the static operation of humanoid robots, then conquer the whole-body collaborative mobile operation, and finally achieve stable dynamic operation capability in open environments with people flow.
For operation accuracy, we continue to iteratively update the control standards in layers, starting from the basic grasping accuracy, gradually polishing to the millimeter level, and finally moving towards sub-millimeter fine operation.
At the same time, we will continue to expand the reserve of standardized atomic actions, master all kinds of practical operations in warehousing scenarios such as picking and placing, waybill pasting, winding and packing, and fully precipitate mature interaction capabilities. Logistics itself is an excellent test field for polishing the above underlying capabilities, and these accumulations can be directly reused when expanding new scenarios outward.
When entering home and elderly care scenarios, we will adopt the implementation idea from simplicity to complexity, first implement low-risk basic tasks such as item delivery and storage, and then gradually iterate the close-range human-machine fine interaction function.
At this stage, when this system enters the home elderly care track, there are still several key problems to be solved.
The first is the insufficient few-shot learning capability. The items and usage requirements in the family environment are highly personalized, and there is an urgent need for stronger in-context learning capability, so that the robot can master new tasks with only a small number of demonstrations.
Secondly, the global safety system needs to be improved. It is necessary to optimize the hardware material protection of the fuselage and battery, build multiple protection mechanisms at the model and control system level, and avoid the risk of bumping and accidental injury in the process of human-robot interaction.
In addition, the adaptive force-tactile control still has room for improvement. The elderly care scenario has extremely high requirements for flexible operation, and the dexterous hand needs to achieve precise, controllable and soft force application. The current dynamic adjustment capability of force perception still needs continuous polishing.
Comments from Investors:
SenseTime Capital stated that logistics is one of the most solid implementation tracks in the current embodied intelligence track, and the value of AI robots in reducing costs and increasing efficiency can be directly quantified; at the same time, compared with the vast majority of entrepreneurial teams with pure technical background, founder Li Qiang has the advantage of industrial cognition in logistics and e-commerce. Combined with the industrial resource synergy of the team and shareholders, the project can quickly enter the core warehousing scenario to run through the real data flywheel, and stand firm before the inflection point of large-scale industrial implementation arrives. The above factors fully conform to the investment logic of the institution "teams that do the right thing in the right scenario".
Yunqi Partners stated that logistics has the characteristics of high-frequency tasks, clear feedback and clear commercial value, and is an important entry point to verify robot capabilities, accumulate real data and improve model generalization capabilities. The progressive route adopted by the Quantum Dynamics team, which takes logistics as the starting point and finally builds a general underlying model oriented to the physical world with cross-scenario generalization capability, is more pragmatic.