95-Postdoc Leads Current Robotics to Secure Series of Early-Stage Funding

The investment community has learned that humanoid robotics company Current Robotics (Yuanliu) has disclosed its first funding round, having completed seed, angel, and Pre-A rounds, with total funding reaching hundreds of millions of yuan.

Investors include renowned market-driven institutions such as Baidu Ventures, Hillhouse Capital, Oasis Capital, Monolith Capital, Qianhai Fangzhou, Fosun Capital, and Junshan Capital, as well as industry players like Zhiyuan, StarMap, and Jike Technology.

Behind the scenes, founder Zhu Yichen, a post-95s entrepreneur, led early research on VLA, publishing over 40 papers at top-tier conferences. Years of technical accumulation have coalesced into a comprehensive infrastructure encompassing full-body dexterous operation primitives, data collection systems, and world model evaluation strategies, continuously translating real human behaviors into humanoid capabilities.

“Humanoid intelligence begins with people.” Following Current Robotics’ philosophy, we are witnessing robots transition from demonstration platforms into homes around the world.

The journey of Current Robotics began with Zhu Yichen.

Born in 1995, he previously served as Head of Embodied Intelligence at Midea Group and was among the first in China to publish research on VLA and world models. He led his team in proposing the first end-to-end VLA paradigm combining diffusion models and VLMs (TinyVLA), extended Chain-of-Thought strategies (DiffusionVLA & DexVLA, published at ICML and CoRL), and was the first to propose using world models for embodied policy evaluation (dWorldEval, published as an ICML Spotlight).

The π0 paper by Physical Intelligence cites only one Chinese team's research—specifically, the work on TinyVLA and ScaleDP completed by Professor Zhu Yichen’s team. Since then, these studies have evolved from VLA models and human demonstration data to whole-body robotic learning, forming the technical foundation of Current Robotics.

Entrepreneurship flows naturally. Yu Chen observed that in the past, humanoid robots either focused on full-body locomotion to improve walking or on hand manipulation to improve grasping, resulting in a scenario where the robot would first walk over and then stop to perform an action. But real-world tasks are rarely this simple—even a simple motion requires stable coordination between the hands and the body.

This requires "loco-dexterous manipulation." Current Robotics' solution is system integration, training a single policy to coordinate whole-body movement, posture adjustment, and fine hand interactions. The resulting loco-dexterous manipulation model, Curr-0, enables robots to continuously adjust their actions while moving. The five publicly demonstrated tasks are all performed by a single policy with shared weights, distinguished by language commands.

The robot carries an object through a doorway and places it in a specific location.

With the model in place, more practical implementation challenges followed—first, data: Current Robotics developed its own full-body exoskeleton data collection system, expanding data collection scenarios from laboratories to more realistic environments such as factories, offices, and homes. Second, faster evaluation and iteration: exploring the use of world models to assess real-world robotic strategies.

To date, the company has focused on Loco-Dexterous Manipulation, developing end-to-end technical capabilities covering data collection, model training, and evaluation and post-training based on world models, aiming to enhance robots’ mobility, manipulation, and generalization across diverse tasks and environments.

This is extremely rare today, given the widespread lack of practicality, data, and generalization in embodied AI. Therefore, Current Robotics attracted simultaneous investment from prominent market-oriented institutions right from its inception, clearly demonstrating strong interest. Even more uncommon is the collaboration among industry players—beyond capital, the greater significance lies in industrial coordination.

The investment community has learned that Current Robotics has joined the Yishu Intelligent Data Ecosystem Alliance, initiated by Xinghai Map, to collaboratively build a complete pipeline from data collection and annotation to application. Additionally, the company is a strategic partner of Mifeng Technology, the embodied intelligence data platform under Zhiyuan.

Investing in capital, orders, and ecosystems is especially prominent in the era of embodied intelligence. After all, only with real-world scenarios can a data flywheel be created, enabling robots to develop the ability to operate stably in the real world at lower costs and in less time.

From the beginning, Current Robotics has focused on filling the gaps in existing embodied AI technology pathways.

For example, to light an incense stick, a natural thought is for the robot to first move to the corresponding position and then stop to perform the action. The corresponding technical approach is to train the movement strategy and the manipulation strategy separately, then integrate them through systems engineering. This approach works in factory assembly lines, but once applied to real-world environments like homes, a noticeable gap emerges between movement and manipulation.

This is not hard to understand: even when a robot remains stationary while completing desk tasks, its waist and lower limbs still need to adjust according to upper body movements. Using only a “stitching” approach, any body tilt disrupts the hand’s force application, causing the robot to appear clumsy.

Current Robotics' Curr-0 whole-body dexterous manipulation baseline was designed specifically for coordination—achieving hand-foot synergy through integrated navigation, full-body balance, and dexterous manipulation. Tasks that are difficult to break down into simple “move” and “do” phases—such as tearing open a tea bag, clearing debris from a table, or carrying an object through a doorway—are executed with impressive proficiency by Current Robotics’ solution.

After the model, the data sources must also be updated synchronously. Mainstream first-person collection can capture a wide range of environments and behaviors, but cannot accurately reconstruct physical interaction information such as human motion states and contact forces; teleoperation-based collection offers higher precision but is more difficult to scale due to equipment and cost constraints.

Can data reflecting environmental, bodily motion, and physical interaction be efficiently collected? To address this, Current Robotics has developed its own full-body data acquisition system, using a head-mounted device, data gloves, a full-body exoskeleton, and electromyography sensors to capture first-person visual input, hand and full-body movements, and whole-body forces in real-world scenarios—all without requiring a robot to be pre-deployed. This approach directly extends data collection into human real-world tasks, offering superior cost-efficiency and applicability while providing the authentic data needed to train humanoid robot brains.

HumanEx Wearable System

Finally, there is the evaluation and post-training phase. Traditional physical robot evaluations rely on robots, specialized facilities, and manual resets, resulting in high costs for hardware, time, and scenario replication. Especially in tasks involving continuously changing contact states, it is difficult to precisely restore the environment to its previous state after a failure, which directly impacts model iteration efficiency.

Based on this, Current Robotics began exploring the use of world models for evaluating robotic policies early on. The team first proposed WorldEval and dWorldEval, which execute robotic policies within generated future worlds and assess policy performance by predicting task progression. Later, in Hi-WM (Human-in-the-World-Model, presented at the RSS 2026 Robot World Models Workshop), they further incorporated human intervention: when the policy encounters errors or uncertainty, humans can directly step in to correct it, and the new interaction trajectories are then used for further training.

The release of CurrentWorld-0 this week extends Current Robotics' prior strategy evaluations and post-training work on world models into a more comprehensive interactive environment. The team defines it as an Interactive World Simulator, integrating cross-ontology, multi-view, and force-tactile prediction into a unified framework for the first time.

Once a robot is replaced with a different body, its action space and control methods also change. Since different robots inherently struggle to share the same low-level control representation, CurrentWorld-0 effectively preserves these differences, allowing each robotic platform to act according to its own characteristics while learning shared environmental dynamics through cross-platform interactions.

After resolving the ontological differences, a new issue arises regarding "how to view." Real robots often rely simultaneously on the head, wrist, and third-person perspectives; the visual output from the same task can vary significantly across different camera angles. However, the underlying object states, robot poses, and task progress must remain consistent across all views. If an object has been moved in one perspective, the other perspectives must synchronize accordingly; even if an object is temporarily occluded, subsequent states must continue from the interactions that have already occurred.

However, when the robot makes actual contact with objects, RGB quickly reveals its limitations. It’s not always possible to directly determine from visual input alone whether the grasp is stable, how forces at the end effector are changing, or whether the object is beginning to slip. CurrentWorld-0 therefore incorporates force and tactile sensing into its predictions, enabling physical feedback that is difficult to capture visually to evolve continuously alongside the robot’s actions.

From folding socks and arranging pillows to slicing cucumbers and opening bottles to pour wine, CurrentWorld-0 incorporates not only rigid body operations but also flexible objects, fluids, and complex contacts into its generative environment. During the policy evaluation phase, the model must continuously respond to the robot’s actual executed actions: if the policy has already misgrasped, subsequent states cannot be restored to common successful trajectories by video priors; failed records that should be exposed must be preserved rather than being “corrected.”

This is also the foundation for using CurrentWorld-0 to compare strategies. Robot strategies can first be run, compared, and tested for failure modes in the simulated environment, with results then validated against real-world performance. The team’s published results show that the relative performance of different policies closely aligns with real robot testing, and the primary failure modes observed in real-world execution can also be reproduced in the simulated environment.

Even after evaluation, these failure experiences can still be leveraged. When a strategy reaches a prone-to-error point, humans can take over midway; the same intermediate state can be saved, rolled back, and then retried with different recovery approaches. The resulting corrective trajectories can further be used for policy post-training, turning issues uncovered in one evaluation into training data for the next round.

At this point, CurrentWorld-0 has established a relatively complete model evaluation and post-training data collection pipeline: strategies are first validated in a generative environment, and once issues are identified, human corrections supplement new interaction data, which is then fed back into post-training.

[CurrentWorld-0 Multimodal Correction Trajectory Result After Post-Training]

As a result, Current Robotics has gradually developed a comprehensive data, model, and evaluation system centered on full-body dexterous manipulation: real-world data provides physical experience, models transform this experience into mobility and manipulation capabilities, and world models further enable strategy validation, failure correction, and post-training to drive continuous improvement in robotic capabilities.

When can the robot start working?

As we can see, embodied intelligence is shedding its storytelling facade and entering the phase of delivering real-world results. No matter how impressive the demo or how dazzling the performance, it must ultimately be applied to real-world task scenarios—far more complex and messy than conceptual models.

On one hand, traditional fixed-station operations only cover a narrow range of physical labor, as most real-world tasks require movement between spaces. On the other hand, robots operate in a continuous, complex, and infinite real world, where the training and iteration efforts are immeasurable, urgently demanding an efficient and streamlined approach to data collection and processing.

From HumanEx and Curr-0 to CurrentWorld-0, Current Robotics is connecting real human experience, robotic skill acquisition, and iterative evaluation to create a full-stack closed loop: front-end collection of real-world behaviors, middle-stage development of full-body dexterous manipulation skills, and subsequent policy validation within a world model, with new data generated near failures flowing back to further training.

The true essence lies in starting with the end in mind.

The market is voting with real money—only those who can truly implement and create tangible industrial value will reap the industry’s rewards. During this period, competition has expanded from building agents to securing data and infrastructure; whoever can establish the earliest data and next-generation model evolution barriers will gain an advantage in the next round of iteration.

In response, Current Robotics provided a definitive solution—assume infinite data and design based on the actual workings of human intelligence. This allows navigation, body balance, and hand movements to be fully coupled, localization to be integrated into an end-to-end system, and data collection to capture richer details within real-world workflows.

The famous closing passage from The Great Gatsby reads: “So we beat on, boats against the current, borne back ceaselessly into the past.” This is the inspiration behind the name Current Robotics—overcoming the limitations of existing technology, swimming against the current, and building a bridge between human experience and humanoid robotics. Just as the company’s logo draws inspiration from Leonardo da Vinci’s Vitruvian Man, symbolizing the human body as a representation of movement, proportion, and scale.

Current Robotics company logo

After all, the source of humanoid intelligence remains human. Through wearable devices that collect data, simulate human movement to create base models, and integrate human-in-the-loop evaluation and post-training, Current Robotics is continuously transforming real human behavior into enhanced humanoid capabilities via a closed loop of hardware, data pipelines, and models.

Current Robotics' wearable data collection system captures signals known as "incidental human behaviors"—natural, free, task-adjacent physical actions that emerge when humans perform real tasks in real-world environments. These physical priors, which are difficult to simulate and cannot be described by verbal instructions, can never fully reveal the unique value of human intelligence in human-robot collaboration without being grounded in original human demonstrations.

Unconscious behaviors are highly important in humanoids. They are sufficiently human-like to generate greater emotional value, often surpassing mere functionality," shared Zhu Yichen. For example, a robot holding objects with both hands or using its body to push open a door—these simple humanoid actions become vivid and engaging.

This scene may not be far off. A vision of harmonious coexistence between humans and machines is unfolding before our eyes—growing from the most authentic human intelligence and ultimately returning to the real world of life.

This article is from the WeChat public account "Investment Daily" (ID: pedaily2012), authored by Yu Mengying.