"Every robotics company is trying to train models that work in the real world, not just in a lab. And the real world is messy. Kitchens don't all look the same. Warehouses have different layouts."
"The way someone chops an onion in a hotel kitchen in New York is different from a catering facility in Houston. There's no standardization on what 'good' video data even looks like yet."
"OpenAI wants one thing, Figure wants another, Boston Dynamics cares about light industrial environments specifically. Some want egocentric video from a camera on someone's head. Others want teleoperation data."
"It's early enough that the whole field is still figuring out the best approaches to gather the necessary data for effective robotics training."
Nabeel Hyatt emphasizes the necessity for diverse and detailed data in robotics to train models that function in real-world environments. Each environment presents unique challenges, such as varying kitchen layouts and different methods of performing tasks. There is currently no standard for what constitutes 'good' video data, as different companies have distinct requirements. The robotics field is still in its early stages, exploring various data types, including egocentric video, teleoperation data, and 3D mapping, to enhance model training.
Read at www.businessinsider.com
Unable to calculate read time
Collection
[
|
...
]