Reliable robot navigation is an active research topic for many real-world applications, such as the automation of industrial equipment, where machines with arbitrary shapes need to navigate very close to obstacles to perform efficiently. Dr Sinyavsky and Dr. Passot, along with Research Engineer Borja Ibarz Gabardos at DeepMind, have developed a new planning architecture that allows wheeled vehicles to navigate safely in cluttered environments.
Their method belongs to the Model Predictive Control (MPC) family of local planning algorithms. It works in the space of two-dimensional occupancy grids and plans in motor command space using a black box forward model for state inference. The method has several properties that make it well-suited for commercial applications: it is deterministic, computationally efficient, runs in constant time, and can be used on platforms of arbitrary shape and drive type
The paper provides a detailed description of the algorithm, showcases its application on real robots, and compares it with other state-of-the-art planning algorithms.