The case for targeted, real-world data in physical AI
Posts discussing Vangrid and Axis Robotics highlight two approaches: using smartphones to gather physical-world context and letting model weaknesses guide what data to collect next.
TLDR
One post argues that bigger models alone may not deliver the next AI breakthrough. It describes Vangrid as exploring smartphone-based collection of real-world context—such as movement, spaces and obstacles—without depending entirely on expensive specialized hardware.
A separate post discussing Axis Robotics advocates a targeted feedback loop: evaluate a model, identify gaps, collect data addressing those weaknesses and retrain. It argues that directing human effort toward those gaps could make each contribution more meaningful, rather than simply adding more examples of familiar patterns.
Combined views
2.2K
2 Sources, first seen ago
