Perceptron AI Launches Isaac 0.5, a Frontier Open-Weight Robotics Model

Article content

Perceptron’s Isaac 0.5 outperforms leading open robot models, including Physical Intelligence’s π0.5 and NVIDIA’s GR00T N1.7

Article content

BELLEVUE, Wash. — 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. Perceptron is working with customers to adapt the model for industrial systems.

Article content

Article content

Isaac can read video, follow language instructions, locate and track objects, estimate the state of a task, and generate robot actions. Industrial automation and robotics teams can use it as the policy that controls a robot or use its visual outputs inside an existing planning and control system.

Article content

Article content

Isaac was trained on three trillion multimodal tokens, one million hours of general video and 100,000 hours of robotics-oriented experience across more than 35 robot systems. Perceptron also established a new scaling law for the data behind robot models. In controlled training experiments, scaling general video from 1,000 hours to one million cut the teleoperation needed to reach the same, well-calibrated action loss from ~5,900 hours to 28.

Article content

Article content

“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,”

Article content

said Armen Aghajanyan, co-founder and CEO of Perceptron AI.

Article content

Frontier performance among open robot models

Article content

Isaac is evaluated on the same kinds of capabilities required in deployed robotics: understanding a scene, following an instruction, adapting to a new task and completing an action sequence. Its results place it at the top of the open-model field across both robot control and embodied reasoning.

Article content

On LIBERO, a standard benchmark for robot manipulation, Isaac averages 97.2% success across spatial, object, goal and long-horizon tasks. The paper reports 97.0% for NVIDIA GR00T N1.7 and 96.9% for π0.5 in the same comparison table.

Article content

Isaac also learns new tasks faster. After one training pass over a single expert demonstration, it reduced error by 7.0x to 10.5x across three unseen tasks. The strongest competing open model, π0.5, improved by 2.3x to 3.1x. GR00T N1.7, MolmoAct2 and SmolVLA also trailed Isaac on every task in the study.

Article content

From an open model to customer deployment

Article content

Perceptron works with customers across manufacturing, logistics, warehousing, security and mobility. The company works directly with teams to adapt Isaac to their cameras, robot hardware, demonstration data and operating workflows.

Article content

Customers can begin with the open model, fine-tune it on their own demonstrations, and work with Perceptron on the final deployment. The same checkpoint can support video analysis, pointing and grounding, task-progress monitoring, and continuous or discrete robot control.

Article content

“A real robot workflow rarely begins and ends with one motion. The system has to understand what it sees, decide what matters and connect that decision to action. Isaac was built to carry that workflow from video and language through to control,”

Article content

Article content

said Akshat Shrivastava, co-founder and CTO of Perceptron AI.

Article content

A new data recipe for robot learning

Article content

Robot demonstrations are expensive because each hour requires hardware, operators and a controlled setup. General video is far easier to collect. While training Isaac 0.5, Perceptron proved a new scaling law for robot learning: video-heavy training mixtures can reduce the teleoperation data requirement by ~210x. This scaling law gives robotics teams a practical way to plan where the next unit of training data should come from.

Article content

Open and available to build on

Article content

Perceptron released both the model and tools needed to evaluate and adapt Isaac for field deployments:

Article content

Article content

Perceptron provides direct support for companies adapting Isaac to commercial deployments.

Article content

About Perceptron AI

Article content

Article content

is an AI research lab pushing the frontier of physical AI, building systems that perceive, reason, and act in the real world, reliably and at scale. Headquartered in Bellevue, Washington, Perceptron was founded in November 2024 by Armen Aghajanyan and Akshat Shrivastava, both former research scientists at Facebook AI Research (FAIR).

Article content

Article content

Article content

Article content

View source version on businesswire.com:

Article content

Article content

Article content

Contacts

Article content

Media contact

Article content

Article content

Article content