Perceptron releases Isaac 0.5: open-source 36B embodied foundation model with null-expert MoE
AkshatS07 · x · 2026-09-11
Perceptron released Isaac 0.5, an open-source embodied foundation model claimed to be the first open model at the frontier of multimodal video understanding, embodied reasoning, and robot control.
- 36B-parameter sparse model using a new "null experts" MoE paradigm: tokens dynamically select expert count by task complexity, giving a 2.5B-model latency profile; visualizations show compute concentrating on text, object borders, and answer-bearing regions
- Trained on 35+ robot systems, 100K hours of robot experience, 1M hours of general video, and 3T multimodal tokens
- Establishes a data-scaling law: scaling general video from 1K to 1M hours cut teleoperation needed for well-calibrated held-out action loss from 5,900 hours to 28
- Can answer video questions, point to and track objects, report task progress, and generate robot actions; weights and paper are open
Related event: Perceptron Open-Sources 36B Embodied Foundation Model Isaac 0.5(2 posts)→
More from Embodied
- Embodied AI startups need a different operational discipline, investor says — ___Mufasaa · 2026-10-27
- π0.5 study: raising demos from 10 to 438 lifts Place Ball in Bowl success from 7% to 83% — stingning · 2026-10-27
- GPT-6 Astra controls dexterous hands zero-shot at just 5.2% success in 500 trials — stingning · 2026-10-27
- HandsOn benchmark: 66 verified tasks and 18,000 human demos for dexterous manipulation — stingning · 2026-10-27
- HandsOn dexterous-manipulation benchmark: best policy 29.1%, GPT-6 Astra just 5.2% — stingning · 2026-10-27
- DexPort: a handheld interface with a real robot hand for dexterous-hand data collection — pabbeel · 2026-10-27