# MI-009: Noninvasive brain-actuated control of a mobile robot by human EEG > 本页由本地论文卡片生成。请不要直接编辑本页；修改源卡片后运行 `vp run docs:generate`。 > 中文支持说明：本页是中文站点镜像，保留论文标题、证据状态和来源字段的原始表述，避免未经核实的翻译改变证据边界。 > 内部 wiki 边界：本页提供论文卡片证据和用户提供的 PDF 下载入口，仅用于私有阅读；全文抽取文本不发布。 > 来源: `library/paper_cards/MI-009.md` ## 论文访问 * 内部 PDF: 下载论文 PDF * DOI / 官方页面: [10.1109/TBME.2004.827086](https://doi.org/10.1109/TBME.2004.827086) * 部署边界: 这些 PDF 链接只适合私有/受保护的内部 wiki。 ## MI-009: Noninvasive brain-actuated control of a mobile robot by human EEG ## Metadata * ID: MI-009 * Title: Noninvasive brain-actuated control of a mobile robot by human EEG * Year: 2004 * DOI / URL: 10.1109/TBME.2004.827086 * Local PDF: 见上方论文访问区块 * Text artifact: local-only path withheld from docs site * Review status: `extracted` ## Study Type * Track: Motor Imagery * Task: online non-invasive EEG control of a mobile robot in a small indoor multi-room environment * Participants or dataset: two healthy volunteers, labeled subject A and subject B; neither had previous meditation or specific mental-training experience * Hardware: commercial EEG cap, Khepera two-wheeled mobile robot, robot-mounted feedback lights, and eight infrared obstacle sensors * Channels or sensors: EEG recorded from F3, F4, C3, Cz, C4, P3, Pz, and P4, referenced to averaged earlobes, sampled at 128 Hz; robot infrared readings were mapped to perceptual states ## Methods * Paradigm: asynchronous EEG BMI maps three mental states to high-level robot commands; the robot combines these commands with perceptual state in a finite-state automaton for forward movement, left/right turns, wall following, and stop behavior * Signal processing or model: surface Laplacian, Welch power spectrum over 8-30 Hz from the last second of data, 96-dimensional feature vectors, statistical classifier with Gaussian prototypes, posterior threshold 0.85, and an `unknown` rejection state; no artifact rejection or correction was used * Training/calibration: subject A trained for 5 days and subject B for 3 days before robot control; daily training used four about-5-minute sessions with 5-10 minute breaks, random task switches every 10-15 s, colored feedback, and offline classifier optimization after each session; both then trained with the robot for 2 days * Online/offline: online BMI responses every 0.5 s, with classifier training/optimization performed offline between sessions; final robot tests compared mental control against manual keypress control using the same robot controller ## Results * Metrics: classifier errors and `unknown` responses below 5% and 30% in mode I; below 2% and 40% in mode II; theoretical channel capacity above 1 b/s in mode I and about 0.85 b/s in mode II; mental/manual operating-time ratio 0.74; average switch between mental commands every 5.0 s * Main findings: two subjects moved a Khepera robot between rooms by mental control only, and the robot never failed to visit the target room at the end of training; performance was worse than manual control but by less than a factor of 1.5 * Reported limitations: only two healthy subjects; miniature mobile robot rather than arm/grasp manipulation; control success depends on robot autonomy and finite-state interpretation of high-level commands; only three mental states were recognized; subtle movement-related EMG could not be fully excluded ## Relevance To This Project * Supports: non-invasive EEG can operate a robot more effectively when decoded states are treated as high-level commands and combined with autonomous robot behaviors * Conflicts with: does not demonstrate robotic-arm grasping, YOLO-based perception, SSVEP target selection, or fine continuous manipulation from EEG * Design implication: MI in SAH-BRI-Grasp should be evaluated as a high-level intervention, mode, or confirmation signal within shared autonomy rather than as a direct low-level actuator command stream ## Extracted Evidence | Claim | Status | Evidence Note | Page/Section | | --- | --- | --- | --- | | Two healthy subjects controlled a mobile robot noninvasively with EEG-recognized mental states. | verified | The abstract and introduction report two subjects moving a robot between rooms by mental control using an EEG BMI that recognized three mental states. | Abstract; Section I | | The robot-control design used high-level mental commands plus robot autonomy rather than direct command-to-motor mapping. | verified | The methods map mental states to high-level commands and use a finite-state automaton plus onboard sensing for wall following, turning, obstacle avoidance, and stopping. | Section II; Figure 2; Section IV | | The EEG setup used eight scalp electrodes and 8-30 Hz spectral features from a surface-Laplacian-transformed signal. | verified | The EEG section lists F3, F4, C3, Cz, C4, P3, Pz, and P4, the 128 Hz sampling rate, surface Laplacian, and 8-30 Hz power-spectrum features. | Section II.A | | The classifier used Gaussian prototypes, a posterior-probability threshold, and an `unknown` response to reduce risky decisions. | verified | The classifier section describes class-conditional Gaussian mixtures, posterior probabilities, a 0.85 threshold, and rejection as `unknown` when confidence is low. | Section II.B | | Training combined participant learning and classifier updates over several days before and during robot operation. | verified | The protocol states that subjects trained for 5 or 3 days, used four short sessions per day, received colored feedback, and then controlled the robot for 2 days while user and BMI adapted to each other. | Section II.C | | Mental control was worse than manual control but achieved an operating-time ratio of 0.74 in the same task/controller comparison. | verified | The results compare mental and manual control for the same room-navigation sequences and report an average operating-time ratio of 0.74. | Section III | | SAH-BRI-Grasp should map MI to high-level shared-control decisions rather than raw low-level robot motion. | inferred | This follows from the paper's conclusion that efficient robot operation required asynchronous high-level commands mapped into a finite-state automaton, not direct motor-action mapping. | Section IV | ## Open Questions * The study has only two healthy participants, so robustness and user variability need stronger evidence before transfer to SAH-BRI-Grasp. * The robot is a miniature mobile platform, not a vision-guided manipulator; grasping and arm-control transfer remains unverified. * The authors argue EEG, not EMG, explains control, but they also state that subtle movements during motor imagery cannot be fully excluded.