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.
- Isaac 0.5 reports a scaling law for combining broad video with robot data.
- At its primary held-out action-loss target, measured teleoperation requirement falls from about 5,900 hours to 28 as general video scales from 1,000 hours to 1M.
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)→
More from Embodied
- Open-source Microduck teaches robots to play football, easing entry to autonomous robotics — tristanbob · 2026-09-04
- Tesla opens Cybercab fleet purchase interest form, signaling third-party robotaxi fleets — JOBhakdi · 2026-09-04
- Open source artificial muscles are coming sooner than you think — IanPritchard · 2026-09-04
- Palmer Luckey hypes mystery silver headset Dime, as investigation digs into OpenAI hardware tease — rohanpaul_ai · 2026-09-04
- $14k Chinese Cars Now Ship With LiDAR, No Subscription — and BYD Covers Crash Losses — aronchick · 2026-09-04
- Losing a badminton match to a robot: a surprising 2026 milestone — lukas_m_ziegler · 2026-09-04