Embodied Artificial Intelligence Reaches Critical Milestone as Independent Variable Robotics Closes Hundreds of Millions of Yuan in Pre-A++ Round Within a Month | Unity Ventures Portfolio News

九合创投·February 17, 2025·1·5

A unified end-to-end Embodied Artificial Intelligence large model is key to improving robots' generalization and adaptability.

Recently, Unity Ventures portfolio company Independent Variable Robotics (X Square Robot) closed a Pre-A++ funding round of several hundred million RMB.

Independent Variable Robotics was founded in December 2023 and is dedicated to developing general-purpose robots through research on embodied artificial intelligence general foundation models. In November 2024, the company announced the completion of a Pre-A and Pre-A+ funding round totaling over 100 million RMB.

Recently, embodied artificial intelligence startup Independent Variable Robotics (X Square Robot) closed a Pre-A++ funding round of several hundred million RMB. This round was led by Luminous Ventures and Legend Capital, with Beijing Robotics Industry Fund and Shenqi Capital participating. The funding will be used for training the next-generation unified embodied artificial intelligence general foundation model and deploying it in real-world scenarios.

The ultimate goal of general-purpose robots is to autonomously execute tasks through interaction, perception, and action — much like humans — with strong generalization and transfer capabilities. The key to achieving this lies in the robot general embodied artificial intelligence foundation model.

Overseas, tech companies including Skild AI, Google DeepMind, and Physical Intelligence (PI) are actively building in this space.

Embodied artificial intelligence can be broadly divided into the "brain" (cognition and decision-making) and the "cerebellum" (motor control). Domestic companies are currently exploring different approaches: some focus on the brain, enhancing robots' language understanding and planning capabilities; others concentrate on the cerebellum, optimizing motor control for walking, grasping, and other movements; and some are pursuing an end-to-end route that unifies both brain and cerebellum — the same path chosen by leading overseas tech companies like Physical Intelligence (PI) and Skild AI.

A unified end-to-end embodied artificial intelligence foundation model is critical to improving robots' generalization and adaptability. Traditional hierarchical architectures can optimize for specific tasks but struggle to adapt to dynamic changes in complex environments. End-to-end approaches enable robots to map directly from perception to motion, forming efficient closed feedback loops that yield stronger autonomous learning and adaptability across multiple tasks and scenarios.

From its inception, Independent Variable Robotics chose the "unified brain-and-cerebellum end-to-end foundation model" route. Founder and CEO Qian Wang told 36Kr that a true embodied artificial intelligence foundation model should use a single model to cover the complete process from perceptual signal input to action output, without artificial layering or modular division. This is the real solution to achieving general embodied artificial intelligence.

Among domestic companies choosing the end-to-end model, technical approaches have also diverged: some prioritize training small models for specific tasks or single scenarios; Independent Variable Robotics, by contrast, adopted multi-task, large-scale scenario training from the start to improve model generality and adaptability.

Wang noted that in the industry today, for complex tasks clearly exceeding single operations, virtually all strong results have been achieved by embodied artificial intelligence foundation models. Small models, which design specific architectures for each task, can typically only execute the most basic single operations without achieving generalization. Foundation models, on the other hand, emphasize engineering approaches to scale up the model until it reaches full generality. The two represent entirely different technology stacks, and accumulating small-model expertise does not effectively transfer to building foundation models.

Generality and generalization are the core defining elements of this generation of embodied artificial intelligence technology. Only by achieving sufficient generality, generalization, and transferability can embodied artificial intelligence truly distinguish itself from traditional automation and enable free operation in unrestricted environments, unconstrained by preset environments or objects.

In November last year, Independent Variable Robotics announced the Great Wall (GW) series WALL-A model — the world's largest-parameter embodied artificial intelligence general manipulation foundation model to date. The model achieves generalization and transfer across diverse physical environment variables and action patterns using extremely few samples, while holding absolute advantages in long-sequence complex operations.

Wang said that after several months of iteration, the WALL-A model now stands at the same level as Skild AI and Physical Intelligence, with some capabilities even surpassing its overseas competitors. In terms of task complexity, it can perform fine operations such as pulling zippers and folding clothes, demonstrating strong adaptability to complex topological structures and physical interactions in stochastic environments. For accuracy on complex tasks, it performs excellently in complex flexible-object manipulation such as folding and hanging clothes, achieving success rates above 90% on tasks lasting several minutes.

Additionally, Independent Variable Robotics' general embodied artificial intelligence foundation model can perform semantic navigation without maps or depth input, make real-time decisions based on video, follow instructions in real time, and conduct autonomous environment exploration.

On the team front, Independent Variable Robotics' core members are based in Shenzhen. Its software and algorithm team has dual backgrounds in robotics learning and foundation models. On the hardware side, the company has assembled core technical leads and executives from leading hardware companies, with mature engineering capabilities and mass production experience.

Founder and CEO Qian Wang holds bachelor's and master's degrees from Tsinghua University and was among the earliest researchers globally to propose attention mechanisms in neural networks. During his PhD, he participated in multiple robotics learning research projects at top U.S. robotics labs, with research experience spanning nearly all areas related to robotic manipulation and home service robots. Co-founder and CTO Allan Wang holds a PhD in computational physics from Peking University and previously served as algorithm lead for the Fengshenbang foundation model team at the International Digital Economy Academy (IDEA), where he led the development of China's first billion-parameter foundation model and one of the earliest trillion-parameter models, Ziya.

Source | 36Kr

Author | Fangyu Wang Editor | Jianxun Su