This project targets a core deployment bottleneck of EEG-based BCIs: user burden from high electrode counts. To preserve wearability while mitigating the spatial information loss inherent to low-channel EEG, the system restores 22-channel EEG from MI-9 sparse observations using a conditional diffusion-based generative model. The restoration is formulated as a conditional inverse problem with an explicit observation mask, and the sampling procedure forcibly preserves observed channels at every reverse diffusion step, ensuring zero distortion on measured electrodes while probabilistically reconstructing missing channels.

On the restored EEG representation, the pipeline performs motor imagery (4-class) intention classification and translates the predicted class sequence into robot motion commands within a ROS2 modular architecture. The design emphasis is system-level: restoration → classification → ROS2 topic publishing is built as a continuous, reproducible chain so that the impact of “sparse-only” versus “restored” EEG inputs can be evaluated consistently at the control interface.

To address real-world safety without replacing user control, the system applies a LiDAR-based, intent-preserving safety filter that performs limited intervention only under hazardous proximity conditions: it decelerates or stops forward motion while preserving rotational intent whenever possible. End-to-end behavior is validated in a Gazebo + ROS2 simulation testbed using parameterized YAML configurations and rosbag-based replay to ensure repeatability of robot responses under identical inputs and environments.