Cloud‑free AI: new chip gives drones and robots a “brain” on board

A research team has introduced a specialized chip designed to run large AI models directly at the network edge – in drones, autonomous robots and industrial controllers. The architecture is optimized for running language and vision models with minimal power consumption, enabling devices to make complex decisions without constant access to cloud servers. For tasks where milliseconds matter – navigating tight spaces or controlling manipulators next to humans – this is far more reliable than sending data to a data center and waiting for a response.

The chip combines energy‑efficient compute blocks for matrix operations, a dedicated accelerator for transformer models and on‑chip memory for storing quantized model parameters. The developers claim that a single compact robot can now run object recognition, trajectory planning and a dialog interface with the operator at the same time. Power consumption remains low enough for integration into battery‑powered platforms, from drones to autonomous carts in warehouses.

Eliminating permanent cloud dependence solves several problems at once: it improves resilience to connectivity loss, lowers operating costs and gives companies more control over where data is stored and processed. Experts believe such chips will become a key driver of “heavy” AI at the edge: smart cameras, robots and drones will not just “see” and “hear,” but also reason locally about situations, adapting to their environment without retraining in a data center. In turn, this will speed up the adoption of autonomous systems in logistics, agriculture, security and urban infrastructure.