Perceptron Releases Isaac 0.5, a 36B Open-Weight Embodied Foundation Model

Perceptron Inc (formerly Meta's Chameleon team) has released Isaac 0.5, an open-weights embodied foundation model. It is a 36-billion-parameter dynamic mixture-of-experts (MoE) model that activates only about 2.5B parameters per token, using a single sparse backbone that unifies multimodal video understanding, embodied reasoning, and robot control in one set of weights. Several authors (@lukasmziegler, @Scobleizer, @rohanpaulai, @codestar) shared or commented on it the same day with consistent information.

Confirmed

• Isaac 0.5 has 36B parameters, a dynamic MoE architecture with 2.5B parameters activated per token, a sparse backbone, and open-weight release
• A single set of weights handles multimodal video QA, object pointing, target tracking, task progress reporting, and robot action generation
• The model can convert web videos into robot training data
• Technical report highlights include the null-expert design (per @codestar)
• The model establishes a scaling law: as general video data scales from 1000 hours to 1 million hours, the teleoperation time needed to reach calibrated action loss drops significantly (per @rohanpaulai)

Why it matters

• Isaac 0.5 replaces costly teleoperation data with video data for training robot control; if the scaling law holds, web videos could become a nearly unlimited source of robot training data, dramatically cutting the data cost of embodied AI
• Open weights let the research community directly reproduce and extend the work—a major open-source advance for embodied foundation models

2026-08-27 ~ 2026-08-27 · 5 related posts

• Isaac 0.5 Released: 36B Dynamic MoE Open-Weight Embodied Foundation Model — code_star · 2026-08-27
• Perceptron Releases Isaac 0.5: 36B Open Weight Embodied Foundation Model — lukas_m_ziegler · 2026-08-27
• [source] NVIDIA releases Isaac 0.5: 36B dynamic MoE open-weight embodied foundation model — Scobleizer · 2026-08-27
• [source] Open-source embodied model Isaac 0.5 establishes video-robot scaling laws — rohanpaul_ai · 2026-08-27
• [source] Open Source Isaac 0.5: 36B Embodied Model Trained on Web Videos — lukas_m_ziegler · 2026-08-27