Perceptron Open-Sources Isaac 0.5: 36B Params for Robots

Perceptron AI has released Isaac 0.5, an open-weight robotics foundation model with 36 billion parameters under Apache 2.0. The launch was reported on 31 August 2026. The model folds video understanding, spatial reasoning and robot control into one sparse backbone, and the company also published code, a technical report and a LeRobot integration.

Key takeaways

36 billion parameters in a sparse MoE architecture — 256 experts per layer, zero to eight routed per token
Trained on 35+ robot systems, 100,000 hours of robot experience, one million hours of general video and three trillion multimodal tokens
A 210x reduction in required teleoperation hours when general video is scaled up
Two control interfaces: a continuous Flow expert with a 36-block diffusion transformer and a discrete 2,048-token FAST vocabulary
Weights on Hugging Face, code on GitHub, Apache 2.0 licence

A scaling law: video instead of teleoperation

Isaac’s strongest result is not a benchmark score but a cost curve. Perceptron trained the model across a grid of data mixtures and measured action-prediction loss on held-out robot trajectories. Out of that came a scaling law tying cheap video to expensive teleoperation.

Hours of general video	Teleoperation required
1,000 h	about 5,900 h
1,000,000 h	about 28 h

210×the drop in required teleoperation once general video is scaled upPerceptron, Isaac 0.5 model card

One backbone, four kinds of output

Isaac takes images, video, language instructions, robot state and action history, and returns text, normalized coordinates, task-state estimates or actions. The base is a 36-billion-parameter Qwen-family backbonebackbone: The model’s main network, shared across all tasks — every output head builds on it. with sparse experts: Mixture of Experts with 256 experts per layer.

Continuous control runs through a dedicated Flow expert and a 36-block Diffusion Transformer, discrete control through a 2,048-token FAST vocabulary. The model predicts the next action chunk while the current one is still executing, closing the control loop in real time.

What the model learns from video alone

Isaac trains on footage without action labels by predicting a “future percept” — a compact description of what will change in the scene. The same network then produces actions.

\dots ▶ Symbol meaning \dots observation history up to time t \dots the future percept: what changed in the scene after Δ \dots the model predicting that percept \dots loss function comparing prediction against reality

Broader scope than the open competition

Perceptron published a comparison of Isaac against other open robot models. The gap is not only in embodiment count but in whether a model was trained for real-time control at all.

Model	Robot embodiments	Real-time-control training	Mistake modeling
Isaac 0.5	35	yes	yes
π0.5	about 7	no	no
SmolVLA	1	no	no
OpenVLA	not reported	no	no
Octo	not reported	no	no

We cover the competing models separately: π0.5, SmolVLA, OpenVLA and Octo. Perceptron was founded — according to The AI Insider — by Armen Aghajanyan and Akshat Shrivastava.

Why it matters

Teleoperation is the cost brake on robotics today: hours of a human on a joystick do not scale the way scraping video does. If Perceptron’s curve transfers to deployments, the entry barrier for building your own policy drops from thousands to dozens of hours of on-robot recording. And the model is fully open-weight under Apache 2.0, with no research-only clauses.

What's next?

Training code, a policy server and evaluation tools are already in the perceptron-ai-inc/isaac repository on GitHub. The checkpoint works only with that repository at a pinned commit:

Bash

git clone https://github.com/perceptron-ai-inc/isaac.git
cd isaac
git checkout be6507b4aed7472f2029606c22684d4ebc9d73e6
git submodule update --init --recursive
cd lerobot
uv sync --locked --extra perceptron_isaac

Stock Transformers and LeRobot are not supported — the model runs only through Perceptron’s repository at the pinned version. The 0.5 label points to further releases, but the company has given no date.