Skild AI has made significant strides in building its foundation model for robotics, and the company is now sharing a sneak peek at its impressive progress. The video showcases the early achievements of Physical AI, a shared brain designed to enable diverse embodiment across various robots, environments, and tasks.

The current state of robotics is hindered by Moravec’s paradox, where hard problems are easy to solve for computers but difficult for humans, while easy tasks are challenging for humans. Skild AI’s approach tackles this issue by developing a general-purpose brain that can handle both simple and complex tasks, such as climbing stairs or assembling fine-grained items.

One of the key features of Skild Brain is its adaptability to different robot morphologies. Whether it’s quadrupeds, humanoids, or table-top arms, the same brain can be adapted to various forms and functions. This versatility is a significant departure from traditional robotics approaches that focus on single tasks or robots.

The video also highlights Skild Brain’s robustness and ability to coexist with humans. The AI model has been trained to apply low forces, making it safe around people. Additionally, it can adapt to disturbances and human interactions, a feature often lacking in other AI models.

Skild AI has tackled the challenge of collecting large-scale robotics data by leveraging simulation and internet videos for pre-training its foundation model. This approach enables the company to achieve scale without breaking the bank on real-world data collection. The team then post-trains this foundation model with targeted real-world data to deliver working solutions.

The company’s focus on diversity in training sets has expanded the available data pool, allowing Skild AI to build a truly physical foundation model that can work across various morphologies. This general-purpose brain is fine-tuned and distilled to meet deployment needs, making it a game-changer for the robotics industry.

Skild AI’s team has been behind several key robotics breakthroughs, including self-supervised learning and cuDNN acceleration. Their innovative approach has opened up new possibilities for robotics research and development. As Skild AI continues to push the boundaries of what is possible with its general-purpose brain, we can expect significant advancements in the field of robotics.

Inspired by

• https://ai-techpark.com/skild-ai-provides-first-look-at-its-general-purpose-robotic-brain/