LimX Dynamics Releases Humanoid Brain System COSA 0.5, Full-Size Robot Oli Completes One-Shot Household Demo

In July 2025, LimX Dynamics released the humanoid brain system LimX COSA 0.5 and demonstrated its full-size humanoid robot Oli autonomously completing multiple household tasks in a single, unedited video. The approximately 3-minute clip, shot in one take without remote control or editing, shows Oli performing tasks such as clothing sorting, toy tidying, box stacking, and deep-bending pickups, all while standing, squatting, and bending—showcasing long-horizon mobile manipulation. This marks the world's second such achievement after Figure, and the first fully autonomous full-size humanoid robot in China to accomplish this.

Technical Architecture: Three-Layer Brain System

COSA 0.5 adopts an S2-S1-S0 three-layer architecture, corresponding to cognition, skills, and control, operating at different time scales in coordination:

• S2 Cognitive Layer (~1Hz): Responsible for scene understanding, memory, reasoning, and task scheduling. It reads visual information via head and wrist cameras to decide "what to do."
• S1 Skill Layer (~50Hz): Contains a set of reusable skills (e.g., VLA), each independently trained and iterated, combined as needed to avoid multi-task interference.
• S0 Control Layer (~1000Hz): Based on the proprietary LimX WBT whole-body motion foundation model (a Transformer policy with tens of millions of parameters, running entirely onboard), it converts high-level motion goals into balanced and coordinated joint commands.

LimX Dynamics emphasizes that models are just skills, while the system is the brain. COSA integrates cognition, memory, scheduling, skills, and whole-body control into a unified system, rather than relying on a single end-to-end model.

Performance Comparison and Industry Positioning

LimX Dynamics compared LimX WBT with the industry's strongest publicly available whole-body tracking model, SONIC:

• Average joint angle error: reduced from 3.3° to 1.5°, a decrease of over 50%.
• Whole-body position error (MPJPE): reduced from 13.75mm to 12.85mm.
• Motion smoothness (joint space jerk): reduced from 129.2 to 115.2.

In long-horizon household task demonstrations, LimX Dynamics is on par with Figure, ahead of Flexion (lower task complexity) and Skild AI (no long-horizon whole-body mobile manipulation shown). LimX believes its differentiation lies in "system completeness" rather than "model parameter scale."

Funding and Open-Source Strategy

In July 2025, LimX Dynamics completed a nearly $200 million Pre-IPO funding round, with a post-money valuation of approximately RMB 15 billion. Investors include IDG Capital and Lens Technology. The company also open-sourced the Humanoid FluxVLA Engine, supporting on-device training and inference for multiple mainstream VLA strategies, aiming to foster an open ecosystem for developer community growth.

Founder's Perspective

Founder Zhang Wei stated in an interview: "There are no technical barriers in the world, only iteration speed." He believes robots already have good brains but lack skills. LimX Dynamics has chosen the path of "brain system + hardware-software integration + open base," contrasting with Figure's full-stack closed-source approach.