Human control of robots over discrete noisy channels with high latency: toward efficient EEG-based brain-robot interfaces

This thesis presents a framework for the design of interfaces that can only obtain noisy and discrete inputs at high latency from a human user (e.g., with an electroencephalograph) to control a robotic system (e.g., a robotic wheelchair) that can provide visual feedback. In this framework, the human user communicates their intent by providing inputs in response to queries posed by the robot. The underlying problem is then to construct a policy that determines the next query to be posed in order for the robot to infer the user’s intent as quickly and as reliably as possible. The approach is to maximize the expected amount of information to be obtained per unit of time from the user’s response given a Bayesian estimate of the user’s intent and an estimate of how quickly and accurately the robot can obtain the user’s response. Under certain conditions, this policy reduces to the optimal feedback policy for transmitting a message between two computational agents over discrete noisy channels. Remarkably, for an interesting class of user intents (e.g., desired paths for robotic navigation), the queries synthesized by the optimal feedback policy can be easily understood and used by humans to convey their intent to the robot. As a case study in the application of this framework, this thesis focuses on the design of EEG-based brain-robot interfaces, which allow human users to control robotic systems with an electroencephalograph (EEG). It presents two interfaces for robotic navigation, where the user’s intent was a desired path to be followed by the robot, and one interface for text entry, where the user’s intent was a desired character to be spelled. The first interface enabled users to navigate a simulated aircraft flying at a fixed speed and altitude over smooth paths that corresponded to a sequence of path primitives. The second interface enabled users to navigate a mobile robot in a virtual indoor environment over paths that (locally) minimized a cost function recovered from human-demonstrated data. These two interfaces provided a new strategy, i.e., navigation based on querying desired paths, which was shown to be advantageous over existing EEG-based interfaces for robotic navigation. The third interface enabled users to specify text commands with inputs obtained from steady-state visually-evoked potentials in EEG at a rate twice as fast as they would using prior state-of-the-art text entry interfaces working with the same input mechanism. This interface showed the importance of querying human intent adaptively based on the prior knowledge on human intent (e.g., likelihoods of characters) and the expected accuracy and latency of inputs