LimX Dynamics Releases LimX COSA, Achieving Brain-Body Integration | Oasis Capital

LimX Dynamics, an Oasis Capital portfolio company, has officially launched LimX COSA (Cognitive OS of Agents). LimX COSA is a newly developed, physical-world-native embodied Agentic OS from LimX Dynamics. By deeply integrating high-level cognition with whole-body motion control, it enables robots to "think, move, and work while thinking" — a unified "big brain + small brain" intelligence operating in the real physical world.

Powered by the COSA system, the full-size humanoid robot Oli becomes the first humanoid agent combining both motor intelligence and high-level cognition.

LimX Dynamics Unveils Its New Embodied Agentic OS: LimX COSA

COSA: A Nervous System That Unites "Knowledge and Action"

As the critical bridge connecting robots to the physical world, LimX COSA is designed as an Agent-centric OS that manages all models, skills, memory, and even emotions — aligning VLA (Vision-Language-Action) with whole-body motion control to achieve systematic integration of "big brain" and "small brain" capabilities, enabling humanoid robots to think proactively and work autonomously in physical environments.

COSA's core design principles:

• Robot-centric architecture, building a modular and reusable multi-skill toolbox where every skill undergoes reliability training and supports independent iteration and compositional calling;
• Unified scheduling, coordination, and planning of skills through multiple foundation models and decision-making mechanisms;
• Achieving "the unity of knowledge and action" — the complete closed loop of understanding tasks, perceiving environments, adjusting decisions, composing skills, and executing actions;
• Serving as an embodied Agentic OS for future humanoid robots, fully leveraging the critical role of intelligent agents.

Full-size humanoid robot Oli awakens for the first time with COSA

As an embodied Agentic OS, COSA adopts a bottom-up three-layer structure, corresponding step-by-step to Oli's behavioral evolution in the video:

• The bottom layer is the cerebellar foundation model (Whole-Body Control Foundation Model), where a highly robust whole-body motion control system provides stable and reliable balance and locomotion;
• The middle layer is a high-level skill layer with big-brain-small-brain fusion, featuring environmental perception and adaptation capabilities that can compose and schedule various skills, supporting complex behaviors such as navigation, obstacle avoidance, grasping, and stair climbing;
• The top layer is autonomous cognition and decision-making, primarily manifested as interaction, memory, and reasoning.

LimX Dynamics' full-size humanoid robot Oli demonstrates mobile manipulation capabilities

If we compare a robot to a person, COSA is essentially its "nervous system." With COSA, the robot is no longer merely a programmed machine, but an embodied intelligent agent with understanding, memory, action, and continuous adaptation capabilities.

Full-size humanoid robot Oli, powered by the COSA system,

becomes the first humanoid agent combining motor intelligence and high-level cognition

COSA: Three Core Capabilities for Embodied Intelligence

• High-level cognition and reasoning: "Big brain" thinking and decision-making based on physical logic

At the top level, COSA endows robots with the logical capacity to proactively understand tasks and goals. Oli no longer relies on pre-programmed code; instead, it can perform natural language semantic understanding and intent recognition. Faced with complex instructions like "bring two bottles of water to the front desk" or "find the red package and deliver it to the workshop," it can autonomously decompose and plan long-horizon tasks, dynamically adjusting task priorities based on environmental changes. This means the robot can handle multiple concurrent tasks around a vague goal in a changing physical environment — truly "thinking while acting."

Full-size humanoid robot Oli with COSA completes mobile manipulation, handing two bottles of water to a guest

• Semantic memory and active perception: Observing the world, accumulating cognition, building a worldview

By constructing cross-temporal, cross-modal perception and memory, COSA gives robots sustained cognitive capabilities in the physical world. Oli can semantically memorize and model environments, objects, and interaction histories, constructing its own "worldview." More importantly, it crosses the threshold from "passively receiving input" to "actively perceiving and exploring" — greeting colleagues, offering condolences when a colleague is sick, connecting and testing printers. The robot knows who is in front of it, where it is, and what it has seen, gathering information through active observation to make more precise judgments.

Full-size humanoid robot Oli with COSA actively checks on a colleague's health

• Real-time perception-based whole-body mobile manipulation: "Big brain + small brain" integrated capabilities

COSA represents the full-stack technology of LimX Dynamics, achieving seamless coupling between high-level intelligence and motor intelligence. Based on real-time perception and whole-body motion control, Oli maintains stable, robust locomotion in complex environments such as continuous long staircases; during movement, it smoothly completes operations like grabbing two bottles of water, picking up trash, and disposing of it — demonstrating remarkable self-adaptation and generalization capabilities. The big-brain-small-brain integration ensures Oli "can think of it, and can do it."

Full-size humanoid robot Oli with COSA autonomously climbs stairs based on real-time perception

COSA: The Watershed Moment From Demo to Product for Embodied Intelligence

COSA is a physical-world-native embodied agent system (Physical-world-native Agentic OS). Its cognition, decision-making, and scheduling mechanisms do not run in abstract computational environments — they take the robot's physical body and the real physical world as their default operating environment. In this architecture, motor capability is no longer merely an output result, but the foundation of robot cognition; robot decision-making is no longer a one-time instruction generation, but a continuous process of real-time interaction and dynamic coupling with environmental constraints. COSA is designed for the genuine deployment of humanoid robots and for creating value for users.

COSA represents a major upgrade for embodied intelligence — shifting focus from model capabilities to OS system capabilities, and from technology demos to product delivery and user experience. LimX Dynamics has taken the industry's first step.