On September 9, ChinaVenture reported that embodied intelligence startup PHYMI had completed a seed funding round of nearly $100 million. IDG Capital led the round, with participation from Yunqi Partners, Didi Chuxing, Fosun RZ Capital, YaoTu Capital and Hesai Technology.

The company said the funding will primarily support core Physical Agent technology R&D, the development of a real-world dynamic data system, product engineering and deployment across use cases, as well as team expansion.

Competition in embodied intelligence continues to intensify. The market's focus is shifting from whether robots can perform isolated actions to whether they can operate in real-world environments, understand users' goals and autonomously complete complex tasks.

From autonomous driving to embodied intelligence

PHYMI founder and CEO Liu Nianqiu previously served as a technical partner and director at DeepRoute.ai, where he helped deploy map-free autonomous-driving systems and advance the company's technology stack toward end-to-end models and vision-language-action, or VLA, models.

Before joining DeepRoute.ai, Liu was a senior engineer at Intel, working on R&D involving new equipment, processes and automation. After leaving Intel, he also participated in several smart hardware projects, gaining experience in hardware-software systems engineering and product development.

Liu said he began following robotics relatively early but did not formally start his own company until 2026. Before then, supply chains for robot hardware, motion control, data and model paradigms were still evolving. The industry remained largely dependent on conventional algorithms and optimization theory, leaving it some distance from AI-driven intelligent systems.

He believes the robotics industry has undergone several important changes in recent years:

First, supply chains have gradually matured, with a growing range of motors and reducers optimized for robots lowering the barriers to manufacturing robot hardware. Second, the availability of AI infrastructure has increased. Third, autonomous driving has established a closed data loop spanning collection, cleaning and annotation, model training, simulation-based validation and over-the-air iteration, offering an approach that embodied intelligence can draw on.

“This means we no longer need to spend a year or two reinventing the wheel. From day one, we can focus on the product itself and build a closed loop in which the product, data and models continuously improve together,” Liu said.

PHYMI was formally established in March 2026.

Physical Agents for the open, dynamic real world

The company's English name is PHYMI. “PHY” stands for the physical world, “M” represents mobility and manipulation, and “I” stands for intelligence.

The company aims to build Physical Agents for the open, dynamic real world, focusing on long-horizon, complex mobility and manipulation tasks.

Liu explained that “long-horizon” refers not only to traveling farther, but also to tasks that take longer, involve more steps, span larger spaces and require more continuous chains of decisions.

Picking up a food delivery, for example, may appear to involve simply moving from one location to another. But in the real world, the item's location, the handoff method and conditions on site can all change. A robot must understand the user's intent, break the goal into tasks and translate them into a series of physical actions. If the environment changes or the user adds a last-minute request, it must adjust its plan and continue until the task is complete.

PHYMI believes many robots today remain largely confined to fixed workstations, predefined processes and specific scenarios, making it difficult for them to handle continuous tasks flexibly in open environments. The company wants its Physical Agents to move and manipulate objects across scenarios, adapting their actions to changing conditions to accomplish complex goals.

Building a full-stack BRAIN, BODY, DATA and LINK system

PHYMI said a Physical Agent is not simply an AI model added to a robot. Models, hardware, data and interaction must work together toward the same goal.

To achieve this, the company has built a full-stack technology system comprising PHYMI BRAIN, PHYMI BODY, PHYMI DATA and PHYMI LINK.

PHYMI BRAIN interprets the environment, breaks down goals and determines actions. PHYMI BODY provides the physical platform that enters and continues operating in the real world. PHYMI DATA captures real-world dynamic data and task feedback across multiple viewpoints and modalities, turning each interaction into cumulative experience. PHYMI LINK connects users, Physical Agents and their environment, carrying goals, permissions, status updates and feedback throughout the task.

Together, the four systems create a closed loop that starts with a goal and spans perception, decision-making, action, feedback and continuous improvement.

The goal: a dependable partner that gets things done

While PHYMI wants to build general-purpose capabilities across scenarios, Liu said the company does not intend to turn its Physical Agents into all-powerful “superhumans,” nor does it aim to replace people entirely.

In his vision, Physical Agents should become dependable, practical and affordable partners for ordinary people, handling frequent, mundane errands such as picking up and delivering items that would otherwise require someone to be physically present.

“It can lend a hand when people are nearby or take charge independently once authorized. It can keep things moving when all goes smoothly, adapt when circumstances change and know when to ask a person for help if it is unsure,” Liu said.

As task and interaction data accumulate, the system will gradually become familiar with its surroundings, learn users' habits and work with them to accomplish goals within the scope of its authorization.

As embodied intelligence moves beyond demonstrations and into real-world applications, PHYMI hopes its Physical Agent technology stack will connect robot hardware, models, data and user interactions, bringing embodied intelligence into a wider range of open, dynamic environments.