LimX Dynamics has demonstrated its full-size humanoid robot Oli autonomously completing a sequence of household tasks in a real home environment without teleoperation or post-production editing1. The 31-degree-of-freedom robot continuously performed clothes folding, item organization, box stacking, trash pickup, and water delivery in a single uninterrupted take.
The demonstration reportedly makes LimX the second company globally, alongside Figure, to achieve autonomous long-horizon household manipulation at this level.
The capability is driven by LimX COSA 0.5, a brain system architecture that the company 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 argued that equating a large model with brain function is misleading, stating that "a pure large model is like Stephen Hawking lying in bed: extremely intelligent but completely unable to move" and that "the brain is not a model; the brain is a system that organizes cognition, skills, and motion control into a coherent architecture running on a real 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 — it decides what to do. The S1 Skill Layer provides a portfolio of trained capabilities including VLA (vision-language-action) models for whole-body motion generation. The S0 Motion Control Layer runs a 10-million-parameter WBT (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.
The Oli demonstration relied on real-robot reinforcement learning with expert teleoperation corrections, independent balance maintenance by S0, corrected data training reward models, and RL iteration on the real robot. LimX says the system gets stronger with continued use rather than freezing after initial training.
LimX says the COSA 0.5 capability unlock makes its consumer-facing strategy increasingly credible. CEO Zhang Wei positions humanoids for commercial, hospitality, entertainment, and household service markets.
ANALYSIS The architectural bet is notable for its divergence from the dominant paradigm in embodied AI, where several leading labs are pursuing end-to-end large models as the primary control system. COSA decomposes cognition, skill execution, and low-level motor control into independently upgradeable tiers with narrow interfaces — the S0 layer's 1,000 Hz on-device loop and the system's modular upgrade path being concrete expressions of that design priority. Whether this layered approach scales to broader task diversity beyond the demonstrated household routines remains an open question.