The New Frontier of AI: Lack of Data for Training Robots

AI giants, including OpenAI, have begun shifting their focus back toward robotics. However, an unexpected obstacle has emerged: there is almost no high-quality dataset available to teach robots how to navigate the physical world. While LLM models evolved thanks to open data on the internet, robots require specialized datasets reflecting physical interactions. This is reported by Techcrunch.com news provides.

To fill this gap, the startup XDOF has launched its operations. The company focuses on collecting, sorting, and systematizing data for robotics. According to TechCrunch, the new project has already raised $70 million in investment from prestigious funds such as Thrive Capital, Spark Capital, and a16z. These funds will be used to create "digital textbooks" for robots.

Data Scarcity and New Solutions

Currently, YouTube videos or simple images used to train robots cannot fully reflect the complexity of the physical world. Philip Wu, founder of XDOF, emphasizes that the next big challenge in AI will not be chips or models, but specifically the lack of data. A perfect robot cannot be created without precise data on how robotic arms grip objects, move them, and navigate various obstacles.

The XDOF team is currently working with nearly 20 clients, including the world's leading AI laboratories. The company not only collects data but also offers cleaning and annotation services. This process helps robots learn from their mistakes and perform physical movements more accurately.

The ABC Project and Academic Collaboration

As a key part of its operations, the company announced a large dataset called ABC in collaboration with UC Berkeley. This collection includes:

• over 130,000 robot manipulation trajectories;
• 300 hours of simulation processes;
• 100 hours of practical evaluation results.

A dataset of this scale was previously unavailable to the scientific community. Such projects are also of great importance for developing countries like Uzbekistan. As the use of robots in manufacturing and services is expected to expand, the existence of such technological infrastructure will help lower robot prices and increase their popularity.

In conclusion, as OpenAI and other major companies return to the robotics race, "data providers" like XDOF will become an integral part of the industry. Physical AI is not just software, but a complex system that operates in harmony with the real world.