Insider Brief
- Perceptron AI released Isaac 0.5, a 36-billion-parameter open-weight embodied foundation model that combines video understanding, reasoning and robot control for applications across manufacturing, logistics, warehousing, security and mobility.
- The model was trained on three trillion multimodal tokens, one million hours of general video and 100,000 hours of robotics experience, with Perceptron reporting that scaling general video reduced the teleoperation needed to reach the same calibrated action loss by about 210-fold.
- Perceptron reported a 97.2% average success rate on the LIBERO robot-manipulation benchmark and has released Isaac’s model weights, technical report and fine-tuning and inference code for developers to adapt the system to their own robots and workflows.
Perceptron AI has released Isaac 0.5, a 36-billion-parameter open-weight model designed to combine video understanding, embodied reasoning and robot control in a single system.
According to the Washington-based company, Isaac can analyze video, follow language instructions, locate and track objects, estimate the state of a task and generate robot actions. Robotics and automation teams can use the model directly as a control policy or integrate its visual outputs into existing planning and control systems.
Perceptron said it is working with customers across manufacturing, logistics, warehousing, security and mobility to adapt Isaac to different cameras, robot hardware, demonstration data and operating workflows.
“Companies need a model that performs at the frontier, learns a new task quickly and adapts to their hardware,” co-founder and CEO Armen Aghajanyan said in the announcement. “Isaac gives them a strong, open starting point, and our team is working alongside our customers to bring it into real operations.”
Isaac 0.5 at a Glance
- Parameters: 36 billion
- Multimodal training: Three trillion tokens
- General video: One million hours
- Robotics experience: 100,000 hours
- Robot systems: More than 35
Perceptron said its training experiments established a scaling relationship showing that larger amounts of general video can reduce the amount of robot teleoperation data needed to reach the same calibrated action loss.
“A real robot workflow rarely begins and ends with one motion,” added co-founder and CTO Akshat Shrivastava. “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.”
In controlled experiments, increasing general video from 1,000 hours to one million hours reduced the amount of teleoperation required from about 5,900 hours to 28 hours, according to the company. Perceptron described that as roughly a 210-fold reduction in the amount of robot-specific data required.
Isaac also posted strong results on robotics benchmarks cited by Perceptron. On LIBERO, a benchmark used to evaluate robot manipulation across spatial, object, goal and long-horizon tasks, Isaac averaged a 97.2% success rate. The company’s comparison table reported a 97.0% success rate for Nvidia GR00T N1.7 and 96.9% for π0.5 on the same benchmark.
Perceptron also tested how quickly Isaac could adapt to unfamiliar tasks. After one training pass over a single expert demonstration, the model reduced error by between sevenfold and 10.5-fold across three unseen tasks, according to the company. In the same comparison, π0.5 improved by between 2.3-fold and 3.1-fold, while GR00T N1.7, MolmoAct2 and SmolVLA also trailed Isaac across the three tasks.
Customers can begin with the open model and fine-tune it using their own robot demonstrations, with Perceptron providing support for adapting the model to field deployments. The same checkpoint can be used for video analysis, pointing and grounding, monitoring task progress and either continuous or discrete robot control, according to the company.
Open for Developers
Perceptron is releasing Isaac 0.5 alongside the tools needed to evaluate, adapt and run the model, giving robotics teams a starting point they can fine-tune using their own demonstrations and hardware.
- Technical report and evaluations: Isaac 0.5: An Open-Weight Embodied Foundation Model
- Model weights: Available through Hugging Face
- Fine-tuning and inference code: Available through GitHub
Perceptron said it also provides direct support to companies adapting Isaac for field deployments. The release allows developers to work from the same model checkpoint for video analysis, grounding, task-progress monitoring and robot control.
