Skild AI: The Universal Brain Revolutionizing Robotics

Skild AI: A Grand Ambition for Robotics

The American start-up Skild AI, valued at $14 billion, stands out for its innovative approach in the field of artificial intelligence applied to robotics. By leveraging simulations and videos of human interactions, the company aims to create a "general-purpose brain" capable of adapting to a multitude of robots, whether humanoid, quadrupedal, or drones. This ambition is based on the idea that the same underlying intelligence can be applied to various types of tasks and environments, whether indoors, outdoors, in factories, hospitals, or university campuses.

Colossal Funding for an Ambitious Vision

Skild AI recently raised $1.4 billion, a sum that reflects investors' confidence in the start-up's potential. Deepak Pathak, co-founder of the company, explains that this financial windfall will enable the development of physical artificial intelligence, seen as the key to achieving artificial general intelligence (AGI). According to him, this advancement could play a crucial role in the reindustrialization of Europe and the United States. Pathak emphasizes the importance of physical AI, which he considers essential for realizing an AGI capable of adapting to various tasks with the same brain.

Omnicorporeal Intelligence: A Revolutionary Concept

Deepak Pathak, at the helm of Skild AI, describes their project as the creation of omnicorporeal intelligence. This general-purpose AI model is designed to operate across a variety of robotic systems, whether robots are functioning in indoor or outdoor environments, in factories, hospitals, or university campuses. The goal is to develop a model capable of adapting to all these contexts, thus offering unprecedented versatility in the field of robotics. This approach aims to overcome the current limitations of robotics, often perceived as a hardware problem rather than an intelligence issue.

The Need for a General-Purpose AI Brain

Traditionally, robotics has been viewed as a hardware-centric field. However, despite impressive technological advancements, robots have yet to become an integral part of our daily lives. Deepak Pathak points out that the real obstacle is not the hardware, but the lack of a general intelligence capable of managing different hardware systems in real-world environments. He insists that what robotics lacks is a brain capable of functioning reliably across various hardware systems, thus allowing for a smoother integration of robots into our daily lives.

Towards a Universal Brain for Robotics

Skild AI's ambition is to create a universal brain for robots, capable of continuously improving through the data generated by each robotic action. This approach, described as a "continuous data-driven improvement loop," allows for the refinement of the foundational model with each deployment, making robots increasingly efficient. Each interaction with the environment provides valuable data that enriches the model, creating a virtuous cycle of continuous improvement that could revolutionize how robots are used across various sectors.

The Challenges of Training Robotic Models

Unlike language models, robotics lacks a vast dataset. This scarcity creates a "chicken-and-egg problem": to collect data, functional robots are needed, but for them to function, data must already exist. Skild AI overcomes this challenge by using videos of human interactions to guide robot behavior, while also relying on simulations for training. Data from teleoperation or virtual reality systems, although limited in quantity, is invaluable for developing robot intelligence.

The Crucial Importance of Simulation

Observation alone is not enough for robots to acquire skills. Just like a tennis player must practice thousands of hours to achieve excellence, robots need to train. Simulations allow robots to practice in virtual environments, thereby reducing the costs of real-world experimentation. By observing human demonstrations, robots can learn to move and interact with their environment before training through millions of simulated interactions, which is essential for honing their skills.

Adapting the Model to Various Applications

Skild AI first develops a general foundation model, the Skild Brain, which they then specialize for specific applications such as industry or logistics. This process generates a data loop that continuously improves the model. One of the major challenges remains the gap between performance in simulation and in real-world conditions, known as the "sim-to-real gap." Skild AI strives to reduce this gap by exposing its systems to a variety of simulated conditions. The goal is to create adaptive systems capable of handling new and unpredictable situations, rather than simply memorizing predefined scenarios.

Understanding World Models and VLA Approaches

World models, often misunderstood, do not aim to create hyper-realistic simulations. Like the human brain, these models reason in an abstract space, capturing the essential properties of the physical world without reconstructing every visual detail. This approach allows for the generalization of skills across different environments and types of robots. Rather than generating realistic internal visualizations, the world model of the future will focus on abstract and efficient reasoning, capable of understanding concepts like force, timing, and balance.

Human-Inspired Learning

Observation-based learning, inspired by cognitive psychology, is central to Skild AI's approach. Robots learn by watching videos and demonstrations, then train extensively through simulation. This shared learning model could enable robots to surpass human capabilities in certain areas. By drawing inspiration from how children learn, robots can understand the intentions and outcomes of interactions, then experiment themselves to refine their skills.

Physical AI as a Pathway to AGI

Deepak Pathak is convinced that physical AI is essential for achieving AGI. Intelligence, in his view, emerges from physical interaction with the world, not from language alone. Language is a manifestation of intelligence, but the deep source lies in our ability to interact with the physical world. Observing evolution, it is evident that intelligence has primarily emerged through physical interactions, reinforcing the idea that physical AI is crucial for developing general intelligence.

Skild AI's Priorities After the Funding Round

With substantial computing resources, Skild AI is focusing on deploying its models across various sectors. Each deployment generates data that enriches the model, ensuring its diversity and effectiveness. Diversity is crucial for building a general model capable of adapting to multiple applications. Skild AI is deploying its systems in industry, warehouses, mobility, and industrial automation, with each sector providing valuable operational data to improve the model.

The Importance of Partnerships and Acquisitions

Robotics requires time and resources to establish itself in each industry. Skild AI is banking on partnerships and acquisitions to accelerate this process. The acquisition of Fetch Robotics, for example, allows Skild AI to leverage Zebra Technologies' experience in deploying autonomous mobile robots, thereby strengthening its position in the sector. These collaborations are essential for entering new markets and developing innovative robotic solutions.