Yunqi Capital Leads Poke Robotics' Nine-Figure Pre-A Round: Tackling the "Manipulation" Challenge in Embodied AI | Yunqi Partners

From "being able to move" to "being able to work" — the core challenge robots still face before truly entering the real world is how to understand physical environments and complete autonomous manipulation in complex tasks.

Recently, Poke Robotics, a Yunqi-Shanghai Jiao Tong University AI Angel Fund portfolio company focused on exploring core embodied intelligence capabilities, completed a nine-figure (USD) Pre-A funding round. As the angel-round investor, Yunqi Capital doubled down in this round. Poke Robotics will focus on robotic manipulation as a key capability, exploring a general embodied intelligence system that integrates World Action Models, reinforcement learning, and real-world data to accelerate the deployment of Physical AI in real-world scenarios. Learn more with Yunqi Capital.

The following is adapted from AI Tech Review. This round was co-led by Shunwei Capital and Matrix Partners China, with participation from Eastern Bell Capital, Jiukun Venture Capital, JunShan Capital, SEE Fund, Liepin Investment, and YuanNuo Capital, among other financial and strategic investors. Existing shareholders including Yunqi Capital, Xiaomi, HLC, Inno Angel Fund, and Orient Capital also continued to invest.

Founded in April 2026, Poke Robotics is still a young robotics company. Among competitors in the home embodied intelligence space, what Poke cares about most is one of the most core and challenging capabilities for embodied intelligence to actually deploy — robotic manipulation, or a robot's ability to work in the real physical world.

Poke Robotics founder Huazhe Xu is an assistant professor and PhD advisor at the Institute for Interdisciplinary Information Sciences, Tsinghua University, and also the founder of TEA Lab, the first lab in China focused on robotic manipulation for embodied intelligence. He received his PhD from UC Berkeley and conducted postdoctoral research at Stanford University.

Within four years of returning to China, Xu has published over 100 high-quality papers in top international journals and conferences including Science Robotics, T-RO, IJRR, NeurIPS, ICLR, CoRL, and ICRA, and received honors including CoRL'23 Best System Paper, RSS'25 Best Paper Finalist, and ICRA'26 Best Paper Finalist. He is one of the few scientists in China with full-stack expertise covering perception, decision-making, control, foundation models, and reinforcement learning for robotic manipulation.

For home embodied robots to go from "being able to move" to "being able to work" — completing highly dexterous, long-horizon complex tasks in non-standardized scenarios like home services — how much road is still left to travel?

Running and jumping don't equal working: "manipulation" is the critical leap

Over the past two decades, the robotics industry has continuously tackled three core problems: finding the way, getting there, and then using hands to complete tasks.

"Navigation" lets robots locate themselves and plan paths. "Locomotion" lets robots maintain balance and move stably across complex terrain. But when a robot walks up to a table yet can't pick up a cup, tidy the table, or cook a meal... no one will pay for embodied intelligence to actually deploy.

To achieve Physical AI — to let a robot's hands interact effectively with the physical world — the key technology is "manipulation," which requires the robot to simultaneously understand object properties, spatial relationships, contact modes, action consequences, and task objectives.

Whether robots can truly enter homes, service, and production scenarios, robotic manipulation may be the most important gate on the path to physical AGI.

But compared to large language models, embodied models have far less room for error: a small mistake in a robot's action can directly cause complete failure in a physical-world task. Therefore, "predicting accurately" is far from enough — understanding the causal relationship between actions and outcomes is the biggest challenge for embodied models.

Poke's approach is not to train separate models for each task, but to build a general embodied intelligence system that understands physical laws, continuously learns new skills, and transfers and reuses capabilities across different tasks.

Rapidly iterating manipulation capabilities: Poke's WAM×RL×DATA flywheel

This June, Poke released a 9-minute fully autonomous real-robot demonstration video. The robot wasn't just picking objects up and putting them down on a table — it cooked a plate of Mapo Tofu from start to finish, on camera.

Don't underestimate this plate of Mapo Tofu. To have a robot autonomously complete the entire dish involves at least five major challenges:

• "Long": Over the nine minutes of cooking, dozens of steps are interlocked, and any tiny error can amplify along the task chain. This is an extreme challenge for the robot's long-horizon planning and full-process state tracking capabilities.
• "Dexterous": Tofu itself is soft, fragile, and deforms under force — not a "friendly" object to manipulate. Cutting tofu requires precise blade placement and exactly the right force, demanding fine force control from the robot approaching human hand levels.
• "Adaptive": As tofu goes from whole block to small pieces, and ground pork goes from raw to cooked, the operating environment changes in real time. The robot must understand what is happening in the pan at each moment and dynamically adjust its actions accordingly.
• "Complex": Throughout the task, the robot must continuously use multiple tools including knives, spatulas, stovetops, and seasoning bottles. After understanding abstract instructions, it must autonomously arrange execution order and actively correct and retry when deviations occur.
• "Precise": Tool handovers, placing items in racks, and similar operations require the robot to control with millimeter-level precision, fitting tools into gaps measured in millimeters.

Beyond Mapo Tofu, Poke also demonstrated folding clothes, threading zip ties, tying sachets, and other high-difficulty manipulations, covering challenges including deformable object handling, fine alignment, and long-horizon tasks.

Globally, previously published cooking real-robot demos have mostly focused on single, short-duration tasks. Achieving fully autonomous, long-horizon manipulation tasks just three months after founding, Poke's secret weapon is a "WAM×RL×DATA" flywheel system.

WAM (World Action Model), RL (full-stack real-robot reinforcement learning), and DATA (high-quality real-world data) form the three core links of Poke's technical system:

Among these three, WAM serves as the foundation for understanding physical causal relationships between actions and environments, responsible for predicting changes in the physical world and planning robot actions;

RL then enables full-stack, large-scale learning through real-robot training environments, reward models, long-horizon action evaluation, and multi-task unified architectures, letting robots continuously learn and evolve through actual interaction;

Meanwhile, Poke's self-developed data collection equipment can mine richly dimensional, high-quality real-world data in home scenarios, driving continuous evolution and iteration of WAM and RL.

Poke has these three components form a mutually reinforcing flywheel, enabling manipulation capabilities to rapidly evolve in the real world.

Defining the true "Poke moment"

To this day, this company less than half a year old has made remarkable progress: from foundation models, real-robot reinforcement learning, data collection hardware to complex task training, Poke has completed full-stack construction. At the same time, Poke has assembled a team covering foundation models, reinforcement learning, Agents, robot bodies, hardware systems, product, and commercialization.

Today, Poke is looking not only at software but also advancing joint R&D of embodied foundation models and robot hardware, exploring scenarios and building ecosystems through software-hardware integration to achieve better embodied intelligence deployment闭环.

When robots can truly enter millions of households, become people's home assistants, and gradually grow into reliable partners that help, serve, and accompany humans — that is when Physical AGI will truly welcome its "Poke moment."

And in Poke's mind, Physical AGI will not suddenly appear one day, but will be born from the compounding returns of the continuous cycle between world models, reinforcement learning, and real-world data — every manipulation, failure, and correction makes the robot understand the world more deeply.