LimX Dynamics Releases Humanoid Brain System COSA 0.5, Completes $200M Pre-IPO Round

In July 2025, general-purpose humanoid robot company LimX Dynamics released version 0.5 of its humanoid brain system, LimX COSA, and completed a nearly $200 million Pre-IPO funding round, achieving a post-money valuation of approximately RMB 15 billion. The company also released a demonstration video of its full-size humanoid robot Oli autonomously completing long-horizon household tasks in a single, unedited, and teleoperation-free take, making it the first full-size humanoid robot in China to achieve such capability.

Technical Breakthrough: Three-Layer Humanoid Brain Architecture

COSA 0.5 adopts an S2-S1-S0 three-layer architecture, corresponding to the cognitive layer (~1 Hz), skill layer (~50 Hz), and motion control layer (~1000 Hz).

• S2 Cognitive Layer: Responsible for scene understanding, memory, reasoning, and task scheduling. It uses visual input from head and wrist cameras along with language commands to decide "what to do."
• S1 Skill Layer: A collection of reusable skills, where the VLA model is just one of them. Different skills are trained and iterated independently to avoid data interference.
• S0 Motion Control Layer: Based on the in-house LimX WBT whole-body motion foundation model (a multi-million-parameter Transformer policy running entirely on-board), it outputs balanced and coordinated joint commands. Compared to the industry's strongest publicly available model, SONIC, joint angle error is reduced from 3.3° to 1.5°, whole-body position error from 13.75 mm to 12.85 mm, and motion smoothness is simultaneously improved.

LimX Dynamics founder Zhang Wei emphasized: "The model is not the brain; the system is the brain." COSA is defined as an embodied agentic operating system, not a single model.

Demonstration Capability: Comparable to Figure, One of Only Two Worldwide

In a continuous three-minute unedited video, Oli sequentially completes tasks including: taking clothes from a high hanger, placing clothes on its arm, accurately throwing them into a laundry basket, storing plush toys, moving and stacking boxes, deep bending to pick up objects, and cleaning up trash. The entire process operates without any human intervention, relying solely on vision and environmental perception for autonomous decision-making.

Currently, only LimX Dynamics and Figure have demonstrated credible long-horizon, uninterrupted, teleoperation-free household tasks globally. Demonstrations from Europe's Flexion and Skild AI lag in task complexity or whole-body motion control stability.

Funding and Commercialization Progress

The Pre-IPO round was participated by IDG Capital, Lens Technology, GGG Group, Redstone VC, Huashan Capital, among others, with overseas capital accounting for approximately 70%. Existing shareholder UAE Leishi Capital has invested in multiple consecutive rounds. Over the past six months, the company has raised a total of $400 million.

Funds will be used for: productization of brain-body fusion technology, batch delivery of thousands of fully autonomous humanoid robots, global market channel development, and overseas manufacturing support. The company has completed share reform and is prioritizing a Hong Kong IPO.

Zhang Wei explicitly stated that none of LimX Dynamics' previous funding rounds included performance-based valuation adjustment clauses, noting that "revenue-based valuation adjustment does not align with the commercial logic of embodied intelligence." The company has secured orders for thousands of units, over half from overseas. The full-size interactive humanoid robot Luna has been delivered to domestic and international customers since May 2025.

Industry Perspective: Demystifying World Models, Emphasizing Skill Data

Zhang Wei believes that the term "world model" is overused in the industry, and what truly matters is the World Action Model for robot decision-making. The biggest misconception in robot intelligence is conflating the brain with skills: the brain is an operating system, while skills require extensive real-world data for training. Hand manipulation remains the biggest bottleneck, primarily dependent on the maturity of tactile sensors and data collection capabilities.