Skild AI Inc., a startup that develops artificial intelligence software for robots, has raised $1.4 billion from a group of prominent investors.

Today’s funding announcement comes about a month after reports of the funding round first emerged. Skild AI has raised about $400 million more than expected. 

The consortium that provided the capital was led by SoftBank Group Corp., which has made robotics a major focus of its AI investment strategy. The OpenAI Group PBC backer recently inked a deal to acquire ABB Ltd.’s robotics business for $5.4 billion. SoftBank was joined in the round by Nvidia Corp.’s NVentures startup fund, Jeff Bezos’ Bezos Expeditions, Salesforce Ventures, Samsung Electronics Co. and more than a half-dozen others.

Skild AI, which is now valued at $14 billion, will use the funding to upgrade its model training infrastructure. The company has developed a foundation model called the Skild Brain that is specifically designed to power robots. The company says that the algorithm has several advantages over competing software.

The training data that developers must use to develop a robot’s AI model varies based on the hardware. An algorithm designed to power robotic arms, for example, must be trained on footage of robotic arms. That complicates development because there is only a limited amount of robotic arm footage available in the open-source ecosystem. The more specialized the robot design, the less training data is readily available.

Skild AI tackled the challenge by equipping its Skild Brain model with a robot-agnostic architecture. The algorithm can run on many different types of robots without the need for extensive fine-tuning. That versatility enables the Skild Brain to learn from videos of humans on the web instead of robot footage.

The company developed the model by creating a simulated environment with more than 100,000 robot form factors. During training, Skild Brain learned to operate each one. Skild AI says the model keeps learning after it’s installed on a real robot. It analyzes data from the host system’s sensors to identify mistakes and correct them.

Under the hood, Skild Brain uses what the company describes as a hierarchical design. One module analyzes the task at hand and generates high-level instructions for the host robot. A second, significantly faster module maps the high-level instructions to specific motion details such as joint angles.

Demo videos on Skild AI’s website shows robots powered by its software watering plants, climbing down stairs and performing a range of other actions. Developers can customize how Skild Brain goes about a task using an application programming interface. The software automatically sorts out implementation details such as how to recover from unexpected setbacks in the middle of a task.

“We believe that a unified, omni-bodied brain is the fastest way to establish a continuous data flywheel where the model gets better with every single deployment, no matter what the hardware or task,” said co-founder and Chief Executive Deepak Pathak.

In addition to enhancing its AI training infrastructure, Skild AI will use its funding round to finance research initiatives. The company plans to explore new model architectures and robot data collection methods.

The funding will also support Skild AI’s commercialization efforts. The company, which recently passed the $30 million annualized revenue mark, is currently working to deploy its software in enterprise environments such as data centers and factories. Skild AI eventually plans to expand its focus to the consumer robotics market.

Image: Skild AI