LimX Dynamics has demonstrated its full-size humanoid robot Oli autonomously completing a continuous sequence of household tasks in a real home environment, without teleoperation or video editing1. The 31-degree-of-freedom robot performed clothes folding, item organization, box stacking, trash pickup, and water delivery in a single uninterrupted take. The company claims the demonstration makes LimX the second company globally, alongside Figure, to achieve this level of autonomous long-horizon household manipulation.
The COSA architecture
The capability is driven by COSA 0.5, a control architecture that LimX describes as "counter-consensus" because it explicitly rejects the prevailing industry approach of treating a single large model as the robot's brain. Founder Zhang Wei argues that equating a large model with brain function is misleading, comparing a pure large model to "Stephen Hawking lying in bed: extremely intelligent but completely unable to move". In Zhang's framing, a brain is not a model but a system that organizes cognition, skills, and motion control into a coherent architecture running on a physical body.
COSA implements a three-tier S2-S1-S0 architecture operating at different time scales. The S2 Cognition Layer functions as a prefrontal cortex analog, using an LLM/VLM agent for scene understanding, memory, world modeling, reasoning, and human interaction to decide what to do. The S1 Skill Layer provides a portfolio of trained capabilities, including vision-language-action models for whole-body motion generation. The S0 Motion Control Layer runs a 10-million-parameter whole-body transformer policy at 1,000 Hz on-device.
The three layers communicate through intentionally narrow interfaces, with intent flowing downward and robot state flowing upward asynchronously without blocking. Each layer can be independently upgraded, replaced, or fine-tuned without affecting the others.
Training on the real robot
The Oli demonstration relied on real-robot reinforcement learning with expert teleoperation corrections, independent balance maintenance by S0, corrected-data training of reward models, and RL iteration on the physical robot. LimX says the system gets stronger with continued use rather than freezing after initial training.
LimX CEO Zhang Wei positions humanoids for commercial, hospitality, entertainment, and household service markets. The company says the COSA 0.5 capability unlock makes its consumer-facing strategy increasingly credible.
ANALYSIS The modular, tiered design contrasts with end-to-end approaches that route perception and action through a single foundation model. By isolating a fast motion-control loop from slower cognitive planning, COSA trades architectural simplicity for the ability to swap components independently, a structure that could accelerate iteration on any single layer without full-system retraining.