# Perceptron AI Releases Isaac 0.5 Open-Weight Robotics Foundation Model

> The 36-billion-parameter model unifies video understanding with embodied control and posts leading results on standard manipulation benchmarks while opening weights and code for enterprise use.

*Published 2026-08-31 · By Diane Okafor*

Isaac 0.5 is a 36-billion-parameter sparse embodied foundation model unifying multimodal video understanding, embodied reasoning, spatial grounding, task-progress estimation, and robot control.

Perceptron AI announced the release of Isaac 0.5 with open weights hosted on the Hugging Face platform and accompanying code on GitHub.

The move supplies enterprises with a starting point for robot learning that can be adapted to proprietary hardware without licensing restrictions.

## What training data supports Isaac 0.5 development?

Training drew from more than 35 distinct robot systems to expose the model to varied mechanical configurations and control interfaces.

One hundred thousand hours of robot experience supplied direct interaction traces while one million hours of general video augmented visual pattern recognition.

Three trillion multimodal tokens completed the pretraining corpus and enabled joint optimization across language, vision, and action spaces.

## How does Isaac 0.5 perform on the LIBERO benchmark?

The model records strong results across multiple task categories designed to test spatial reasoning, object handling, goal specification, and extended sequences.

These outcomes position the open model ahead of several closed and open alternatives previously evaluated on the same suite.

## What technical capabilities are unified in the architecture?

The sparse design integrates video encoding, spatial grounding, progress estimation, and low-level control within a single parameter set.

This unification reduces the need for separate specialist models and simplifies deployment pipelines in production environments.

Training data composition for Isaac 0.5Data CategoryQuantityContribution to ModelRobot systemsMore than 35Diverse hardware exposureRobot experience100,000 hoursEmbodied interaction dataGeneral video1,000,000 hoursMultimodal video understandingMultimodal tokens3 trillionLarge-scale pretraining

## Where can developers obtain the model and code?

Weights reside at the Hugging Face repository PerceptronAI/Isaac-0.5.

Fine-tuning scripts and inference utilities appear in the perceptron-ai-inc/isaac GitHub repository under the Apache 2.0 license.

These artifacts allow immediate download, modification, and on-premise execution.

## What market implications arise for robotics stakeholders?

Open weights lower barriers for companies that must customize policies to unique factory floors or warehouse layouts.

Hardware vendors can integrate the model without negotiating per-seat fees common in closed ecosystems.

- Download weights from Hugging Face to begin local evaluation.
- Clone the GitHub repository and review Apache 2.0 terms for compliance.
- Run initial tests on internal robot fleets using LIBERO-style task suites.
- Engage Perceptron AI support for hardware-specific fine-tuning projects.

## How have industry experts responded to the announcement?

Company leadership emphasized the practical value of a frontier open model that adapts quickly to customer hardware.

> Companies need a model that performs at the frontier, learns a new task quickly and adapts to their hardware. Isaac gives them a strong, open starting point, and our team is working alongside our customers to bring it into real operations.Armen Aghajanyan, co-founder and CEO of Perceptron AI

## What developments are anticipated in the coming months?

Perceptron AI indicated ongoing collaboration with early customers to validate performance in live operations.

Further releases may include additional fine-tuned variants targeting specific robot morphologies or industry verticals.

Community contributions to the GitHub repository are expected to accelerate once the codebase stabilizes.

## Sources

1. [Isaac 0.5 brings multimodal video understanding, embodied reasoning, spatial grounding, task-progress estimation, and robot control into one 36-billion-parameter sparse model.](https://huggingface.co/PerceptronAI/Isaac-0.5)
2. [Perceptron AI has launched Isaac 0.5, a 36-billion-parameter open-weight embodied foundation model that combines video understanding, embodied reasoning and robot control, the first open model at the frontier of all three.](https://www.businesswire.com/news/home/20260828535061/en/Perceptron-AI-Launches-Isaac-0.5-a-Frontier-Open-Weight-Robotics-Model)
3. [We trained Isaac on data from more than 35 robot systems, 100,000 hours of robot experience, one million hours of general video, and three trillion multimodal tokens.](https://github.com/perceptron-ai-inc/isaac)
4. [Perceptron AI released Isaac 0.5 with open weights available on Hugging Face, including fine-tuning and inference code on GitHub.](https://superpowerdaily.com/research/release-velocity)

---
Source: https://aiintelreport.com/enterprise-ai/perceptron-ai-isaac-0-5-open-weights-robotics
Index: https://aiintelreport.com/llms.txt · Full text: https://aiintelreport.com/llms-full.txt