Perceptron’s Mk1.5 Model Beats Egocentric AI at 25x Lower Cost

A new intelligence layer for robots launched by Perceptron allows for the zero-shot control of quadrupeds and drones without requiring retraining for specific platforms. The model, known as Mk1.5, reduces latency by 2x to 5x compared to its predecessor and establishes a new state of the art in egocentric understanding, outperforming competing models at 25x lower cost. It incorporates audio as a primary modality, enabling capabilities such as predicting an object's proximity using the Doppler effect. To improve efficiency, Perceptron separated reasoning from world knowledge by integrating external tool calls, such as search and reverse image search. The system can also spawn parallel sub-agents to decompose complex tasks, such as identifying specific food items in a basket to check for allergies. This release is the first step in the company's plan to build agentic foundation models for real-time physical world applications.