Archive — US — techcrunch
Collecting robot training data is dirty, unglamorous work. Some AI labs arealready paying XDOF to do it.
OpenAI is relaunching its robotics program, emphasizing the need for physical training data, which is currently lacking. Startup XDOF aims to bridge this gap by creating data pipelines and collection systems, raising $70 million to support its mission. With a partnership with UC Berkeley, XDOF will release extensive robot training data, enhancing capabilities in the AI community as physical AI becomes the next frontier in technology.
Key points
- OpenAI's reboot of its robotics program points to the need for reliable training data. XDOF was founded to address this gap by developing infrastructure for data collection and cleaning.
- XDOF raised $70 million to create data systems, focusing on teleoperation data and partnerships for high-quality training data.
- The company plans to collect egocentric data through wearable sensors and teleoperation, leveraging human operators to gather complex data.
- Existing data sources like YouTube are inadequate for robotics, necessitating a specialized approach to training data for machine learning models.
Why it matters
- The emergence of XDOF highlights a broader trend in AI towards specialized data infrastructure, similar to how data scarcity initially hindered language model development.
- The robotics sector is racing to catch up with the more mature AI fields, reflecting the challenges faced by tech companies in adapting to new frontiers.
- As physical AI becomes increasingly critical, the ability to generate quality training datasets may define the next wave of technological advancements in the industry.
XDOF's efforts underscore the critical need for specialized data solutions in robotics, revealing how foundational training data will shape the future of physical AI, paralleling previous developments in AI and machine learning.