Perceptron Releases Open-Source Embodied AI Model Isaac 0.5

Episode 1 · Perceptron releases Isaac 0.5, a 36B open-weight embodied foundation model (2026-08-27, 10 posts)

Perceptron Inc (the former Meta Chameleon team) released Isaac 0.5, an open-weight embodied foundation model. It is a 36B-parameter dynamic MoE model activating only 2.5B parameters per token, using a single sparse backbone to unify multimodal video understanding, embodied reasoning, and robot control. Multiple authors (@lukasmziegler, @rohanpaulai, @codestar, @jparkerholder, @AkshatS07) covered the release consistently on the same day; @Scobleizer mistakenly attributed it to NVIDIA, while all other posts confirm Perceptron as the publisher.

Confirmed

• Isaac 0.5 is a 36B dynamic MoE model with 2.5B active parameters per token, a sparse backbone, and open weights
• A single set of weights handles video QA, object pointing, goal tracking, task-progress reporting, and robot action generation
• The model converts web video into robot training data and was pretrained on action-free video
• The technical report highlights designs such as null-experts (per @codestar)
• The model establishes a scaling law: as general video data grows from 1,000 to 1 million hours, the teleoperation time needed to reach calibrated action loss drops from 5,900 hours to 28 hours (per @rohanpaulai and @AkshatS07)
• @rohanpaulai adds that Perceptron combined massive cheap video with expensive robot demonstrations; with enough video, action-learning quality matches that of large teleoperation datasets

Why it matters

• Investor David Cowan (via @AkshatS07) notes that action-data collection is one of robotics' biggest costs; Isaac 0.5 replaces costly teleoperation data with video, cutting required teleoperation from 5,900 to 28 hours (210x) and making web video a near-unlimited source of robot training data
• Open weights let the research community reproduce and extend the work, marking a significant open-source advance for embodied foundation models

• 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
• NVIDIA releases Isaac 0.5: 36B dynamic MoE open-weight embodied foundation model — Scobleizer · 2026-08-27
• Open-source embodied model Isaac 0.5 establishes video-robot scaling laws — rohanpaul_ai · 2026-08-27
• Perceptron trains Isaac on massive video data to reduce robot training needs — rohanpaul_ai · 2026-08-27
• Open Source Isaac 0.5: 36B Embodied Model Trained on Web Videos — lukas_m_ziegler · 2026-08-27
• Isaac 0.5 scaling law: teleoperation need drops from 5,900 to 28 hours as video scales — AkshatS07 · 2026-08-27
• Perceptron releases Isaac 0.5, a 36B embodied foundation model — AkshatS07 · 2026-08-27
• Perceptron releases Isaac 0.5: An embodied foundation model cutting teleoperation needs by 210X — rohanpaul_ai · 2026-08-27
• Isaac 0.5 released: 36B open-weight MoE foundation model for robotics — jparkerholder · 2026-08-27

Episode 2 · Perceptron Releases Open-Weights Embodied Model Isaac 0.5 (2026-09-03, 2 posts)

Perceptron has released Isaac 0.5, a 36B open-weights embodied foundation model spanning video understanding, reasoning and robot control. Demonstrations like T-shirt folding transfer across 35 robot embodiments.

• Isaac 0.5: open-source 36B embodied foundation model trained on 1M hours of video, weights released — AkshatS07 · 2026-09-03
• Perceptron's Isaac 0.5 robot folds t-shirts, ships open weights for cross-embodiment repro — iamrobotbear · 2026-09-04