Axis Robotics shifts to targeted manipulation data

Brute-force data scaling worked for language models, but physical AI cannot simply copy that playbook without hitting steep diminishing returns.

* Raw trajectory volume often introduces repetitive, low-signal noise that inflates training costs without addressing critical physical edge cases.

* Targeted data engines leveraging crowd-teleoperation and verifiable simulation pipelines provide richer distributions for robust generalization.

* The competitive moat in physical AI is moving away from data hoarding toward curating task-specific datasets that directly solve real-world manipulation failures.