Isaac 0.5 scaling law: teleoperation need drops from 5,900 to 28 hours as video scales

AkshatS07 · x · 2026-08-27

Highlighting Perceptron's Isaac 0.5 (36B params, 2.5B active per token, one sparse backbone for video perception, embodied reasoning and robot control), investor David Cowan pointed to its key economic implication: action-data collection is one of robotics' hardest costs.

Cheap large-scale video could nearly substitute for expensive teleoperated robot data, potentially rewriting the cost structure of robot data collection.

Related event: Perceptron releases Isaac 0.5, a 36B open-weight embodied foundation model(10 posts)→

Original post →

More from Embodied

Embodied channel →