Isaac 0.5

Isaac 0.5 is a model from Perceptron, a research company building models for physical AI. It is described as a promptable vision model that takes images, video and language, and, according to a YouTube creator's summary, also robot state, to answer questions or predict robot actions. The same summary says it was trained on 35+ robot systems, about 100k hours of robot data and roughly a million hours of video. That figure comes from a third-party video and was not confirmed on the provided official pages.\n\nPerceptron's site highlights advanced spatial perception, physical reasoning (cause and effect, object dynamics) and real-time operation for time-critical uses. One example on the site asks the model to judge whether a basketball shot was taken before the buzzer, using ball position and shot clock, and to give a one-sentence rationale.\n\nAccess is through Perceptron's APIs, or through a commercial license to the weights on request. The terms of use describe a hosted service with consumer and enterprise licenses and limits on reverse engineering and derivative works. One creator called it open source, but the official pages describe only commercial weight licensing, so open weights are not confirmed here.