Perceptron AI Launches Open-Weight Robotics Model Called ‘Isaac 0.5’
aiinsider ·
Perceptron AI has released Isaac 0.5, a 36-billion-parameter open-weight embodied foundation model that combines video understanding, reasoning, and robot control. The company says the model can interpret video, follow language instructions, track objects, assess task states, and generate robot actions for uses including manufacturing, logistics, warehousing, security, and mobility. Isaac was trained on three trillion multimodal tokens, one million hours of general video, and 100,000 hours of robotics experience across more than 35 robot systems. In controlled experiments, Perceptron reported that increasing general video training from 1,000 to one million hours reduced required teleoperation data from about 5,900 hours to 28 hours, or roughly 210-fold. The model achieved a reported 97.2% average success rate on the LIBERO manipulation benchmark and showed sevenfold to 10.5-fold error reductions on three unfamiliar tasks after one pass over a single expert demonstration. Perceptron has released the model weights through Hugging Face, along with a technical report and fine-tuning and inference code on GitHub, while offering support for companies adapting Isaac to their robots and operational workflows.
aiinsider