# BRI-004: An Augmented Reality Brain-Robot Interface for Generalist Robot Arm Manipulation > Generated from a local paper card. Do not edit this page directly; edit the source card and rerun `vp run docs:generate`. > Internal wiki boundary: paper-card evidence and user-provided PDF downloads are available for private reading. Extracted full-paper text is not published. > Source: `library/paper_cards/BRI-004.md` ## Paper Access * Internal PDF: Download Paper * DOI / official page: [needs confirmation](https://arxiv.org/abs/2606.16413) * Open-access page: [Open access page](https://arxiv.org/abs/2606.16413) * Deployment boundary: these PDF links are intended only for a private/protected internal wiki. ## BRI-004: An Augmented Reality Brain-Robot Interface for Generalist Robot Arm Manipulation ## Metadata * ID: BRI-004 * Title: An Augmented Reality Brain-Robot Interface for Generalist Robot Arm Manipulation * Year: 2026 * DOI / URL: https://arxiv.org/abs/2606.16413 * Local PDF: see Paper Access section above * Text artifact: local-only path withheld from docs site * Review status: `extracted` ## Study Type * Track: AR BRI / Shared Autonomy / Robot Manipulation * Task: augmented-reality brain-robot interface for generalist robot-arm manipulation in multi-step ADL-like tasks * Participants or dataset: 18 healthy participants, three tasks: drinking, drawer use, and oven operation * Hardware: Meta Quest Pro AR headset, g.tec g.GAMMAcap2 EEG, Franka Emika Panda arm, distributed Python/SocketIO modules * Channels or sensors: 22-channel EEG at 256 Hz; headset eye tracking at 72 Hz; headset passthrough images for object detection; robot cameras and proprioception for the VLA policy ## Methods * Paradigm: gaze-based object selection plus MI-based action selection in a shared-autonomy framework * Signal processing or model: OWLv2 zero-shot object detection for fixated objects; MI decoding with 8-32 Hz bandpass, FBCSP features, and RBF SVM; fine-tuned pi0.5 VLA policy for robot execution with OSCBF safety constraints * Training/calibration: per-participant MI calibration with 48 trials; robot policy fine-tuned on about 90 minutes of teleoperation data across nine tasks * Online/offline: online feasibility study with real robot task execution ## Results * Metrics: subtask success, task completion time, offline and online EEG decoder accuracy, SUS, NASA-TLX * Main findings: drawer and oven tasks achieved 100% subtask success; place-mug subtask succeeded in 18/20 attempts after two resend events; offline MI test accuracy averaged 0.70 +/- 0.17 and online decoding accuracy averaged 0.86 +/- 0.23; SUS was 76.94, in the Good range * Reported limitations: healthy participants only; separate EEG cap and AR headset caused wearability challenges; EEG/robot setup still requires technical expertise; target population with motor impairments was not evaluated ## Relevance To This Project * Supports: a close noninvasive AR BRI comparator using gaze for object selection, MI for action selection, and shared autonomy for robot manipulation * Conflicts with: uses gaze plus MI rather than SSVEP plus MI; uses VLA policies rather than a YOLO-to-grasp pipeline; currently a 2026 arXiv preprint with DOI/status needing confirmation * Design implication: SAH-BRI-Grasp should explicitly compare its SSVEP scene-command-space design against gaze/AR BRI approaches and should report usability, workload, decoding accuracy, and task success together ## Extracted Evidence | Claim | Status | Evidence Note | Page/Section | | --- | --- | --- | --- | | The system combines gaze-based object selection and MI-based action selection. | verified | The abstract, introduction, and system sections describe eye tracking for targeting and MI for Place/Use action control. | Abstract; Sections I, III.B-III.D | | The interface is a shared-autonomy BRI where the user specifies high-level intent and the robot executes manipulation policies. | verified | The background and robot-control sections describe turn-taking shared autonomy and VLA policy execution. | Sections II.B, III.E | | The feasibility study used 18 healthy participants on three multi-step ADL-inspired tasks. | verified | The experiment section lists participant count and the drinking, drawer, and oven tasks. | Sections IV.A-IV.B | | Online EEG decoding and error-recovery mechanisms improved effective action selection beyond raw offline accuracy. | verified | Results report offline test accuracy of 0.70 +/- 0.17 and online decoding accuracy of 0.86 +/- 0.23, attributing improvement to sliding-window decoding and gaze-based recovery. | Section V.A | | BRI-004 is a strong comparator for SAH-BRI-Grasp's shared-autonomy and usability framing. | inferred | The paper is noninvasive and robot-manipulation oriented, but its input split, visual model, robot policy, and tasks differ from SAH-BRI-Grasp. | Full paper | | Generalization to assistive target users is not verified. | needs confirmation | The authors state all participants were healthy adults and target users with motor impairments were not evaluated. | Section VI | ## Open Questions * Should SAH-BRI-Grasp include a gaze-only or gaze-plus-MI baseline, or only cite BRI-004 as external comparator? * Which workload/usability instruments should be mirrored: SUS, NASA-TLX, or a shorter task-specific questionnaire? * How should the paper distinguish dynamic SSVEP command spaces from gaze dwell object selection?