Brain-Robot Interface

• Brain-robot interface is a multidisciplinary system that integrates EEG decoding, robot control, and interactive frameworks for translating neural intent into robotic action.
• It employs varied EEG paradigms like MI, SSVEP, and P300, utilizing both traditional signal processing and deep learning methods to optimize decoding accuracy.
• These systems leverage shared autonomy and multimodal control to enable practical applications in rehabilitation, telepresence, and domestic assistance.

Brain-robot interface (BRI) denotes systems in which brain activity is acquired and interpreted to influence the behavior of a physical robot, either passively or actively. In the recent EEG-centered literature, BRI is commonly organized around three entities—Brain, Robot, and Interaction—with the interaction layer mediating both command flow from the user to the machine and feedback from the machine back to the user (Zhang et al., 2024). Within this framework, BRIs range from noninvasive social-robot gesture control driven by imagined kinematics (Abiri et al., 2017), to telepresence navigation with ROS-based shared control (Beraldo et al., 2017), to assistive manipulation with P300 and inverse kinematics (Arrichiello et al., 2019), to general-purpose household robotics supported by modular EEG decoding and robot learning (Zhang et al., 2023), to augmented-reality and multi-brain multi-robot systems that couple neural decoding with edge AI, visual grounding, and shared autonomy (Ouyang et al., 2024, Zhang et al., 15 Jun 2026).

1. System model and interaction structure

A recent review of 87 studies published during 2018–2023 characterizes EEG-based BRI through the three-entity model of Brain, Robot, and Interaction (Zhang et al., 2024). The Brain entity includes signal acquisition and signal decoding; the Robot entity spans industrial, service, medical, social, educational, exploratory robots, and autonomous vehicles; and the Interaction entity specifies how neural signals are translated into control or modulation of robot behavior.

Entity	Main dimensions	Representative emphasis
Brain	Signal acquisition, signal decoding	EEG paradigm, device, electrode count, preprocessing, AI methods
Interaction	Pure BCI, BCI + Agents, Proactive Control, Task-oriented HRI	Direct control, hybrid signal control, AR/VR/XR, shared tasks
Robot	Multiple robot categories	Manipulation, mobility, social interaction, assistance

The review further distinguishes Pure BCI from BCI + Agents, and separates Proactive Control from Task-oriented HRI (Zhang et al., 2024). This classification is important because many practical BRIs are not simple one-way “brain-to-actuator” channels. Instead, they are composite systems in which neural input is combined with robot autonomy, computer vision, state machines, map-based navigation, or AR overlays. Systems such as BRIEDGE explicitly include sensing, computing, communication, and control as a single end-to-end architecture for multi-brain to multi-robot interaction (Ouyang et al., 2024), while AR-based manipulation systems use eye tracking, EEG, object detection, and robot policy execution within one loop (Zhang et al., 15 Jun 2026).

A recurrent structural feature is shared autonomy. In telepresence navigation, the user issues only left/right commands while the robot proceeds forward autonomously and avoids obstacles (Beraldo et al., 2017). In NOIR and NOIR 2.0, the human specifies high-level intent—“what,” “how,” and “where”—and the robot executes parameterized primitive skills (Zhang et al., 2023, Kim et al., 25 Nov 2025). This suggests that contemporary BRI is best understood as a family of layered cyber-physical systems rather than a single decoding problem.

2. Neural paradigms, signal processing, and decoding

The dominant noninvasive substrate is EEG. Across the surveyed literature, common paradigms include motor imagery (MI), steady-state visually evoked potentials (SSVEP), P300, ERP, and ErrP, with traditional pipelines relying on filtering, artifact removal, CSP, CCA, LDA, SVM, Hjorth parameters, and Welch-PSD, and newer systems incorporating CNNs, LSTMs, GRUs, and hybrid architectures (Zhang et al., 2024).

BRI decoding strategies vary substantially by task. One line of work decodes continuous imagined kinematics rather than discrete classes. In a social-robot system using an Emotiv EPOC headset with 14 channels, user-intended cursor velocities are estimated from time-domain EEG by linear regression with five time lags:

$$u[t] = a_0^x + \sum_{n=1}^N \sum_{k=0}^K b_{nk}^x e_n[t-k], \qquad v[t] = a_0^y + \sum_{n=1}^N \sum_{k=0}^K b_{nk}^y e_n[t-k],$$

with $N=14$ and $K=5$ (Abiri et al., 2017). The decoded cursor position is then thresholded into left- or right-hand robot gestures.

A second line relies on frequency-tagged selection. In NOIR, objects flicker at unique frequencies and are classified by canonical correlation analysis, using sine/cosine reference templates and choosing the frequency with the highest correlation (Zhang et al., 2023). NOIR 2.0 retains SSVEP for object selection, with frequencies such as $6, 7.5, 8.57, 10\,Hz$, and uses CCA within a modular “what/how/where” pipeline (Kim et al., 25 Nov 2025). A separate humanoid navigation system replaces fixed GUI boxes with variable natural-scene SSVEP stimuli: object regions detected by SSD in the NAO camera feed are flickered at $10$, $12$, and $15\,Hz$, and a 1D CNN termed the SSVEP Convolutional Unit performs the EEG classification (Aznan et al., 2018).

Other systems prioritize event-related or menu-driven selection. A P300-based assistive manipulator uses a multi-layered BCI2000 GUI with flashing icons, SWLDA-based classification, and UDP/IP transmission of high-level commands to ROS, where those commands trigger perception and kinematic control (Arrichiello et al., 2019). Low-cost artifact-centered approaches also exist: a Raspberry Pi/PiEEG platform detects voluntary blinks by bandpass filtering and FFT-based thresholding in low-frequency ranges, then maps those detections to GPIO outputs for LEDs or a toy robot (Rakhmatulin et al., 2022).

Deep architectures appear when personalization or multi-class decoding is emphasized. DeepBrain uses a low-cost Brainlink device with two dry electrodes and classifies four states—focused, relaxed, focused→relaxed, relaxed→focused—via dense layers, stacked LSTMs, and an attention selector, trained with Adam and crossentropy (Wu et al., 2020). A hybrid MI/SI system for robotic arm control employs a hierarchical teacher CNN with separate MI and SI branches, then transfers knowledge to a simpler student model through distillation loss and soft labels (Lee et al., 2022). BRIEDGE introduces an EEG-adaptive neural network with intermittent masking and informer-based ProbSparse self-attention for heterogeneous EEG sources, coupled to a semantic encoding-decoding communication pipeline (Ouyang et al., 2024).

The decoding layer is often hybridized with non-EEG channels. NOIR and NOIR 2.0 use jaw clench or facial muscle signals as confirmation and interruption mechanisms (Zhang et al., 2023, Kim et al., 25 Nov 2025). A soft-robot system uses only three bipolar EEG channels and a two-stage LDA scheme in which jaw clench selects the active axis and right-hand MI determines movement sign (Stölzle et al., 2024). This diversity of pipelines indicates that BRI decoding is highly task-specific, with continuous regression, discrete selection, hybrid confirmation, and semantic compression all coexisting within the field.

3. Robot control, autonomy, and command execution

Decoded neural intent must be transformed into robot behavior through a control architecture. In many BRIs, this is the decisive engineering layer.

For mobile robots, shared navigation is a common solution. A telepresence system integrating a non-invasive SMR-based BCI with ROS sends left/right commands via UDP to a Pepper robot; ROS then handles localization, mapping, obstacle avoidance, and recovery behavior. The robot exploits one map for localization and one for navigation, and proceeds forward by default while accepting only path-feasible BCI turns (Beraldo et al., 2017). The probability of user intention is temporally integrated through exponential smoothing,

$p(y_t)=\alpha p(y_{t-1})+(1-\alpha)p(y_t|x_t),$

to stabilize command emission (Beraldo et al., 2017).

Manipulation-oriented BRIs often expose only high-level actions to the user. In the P300-driven assistive robot, user selections such as “Move water bottle right” are transmitted to ROS, which activates perception and a closed-loop inverse kinematic controller that manages multiple equality-based and set-based tasks simultaneously (Arrichiello et al., 2019). The control law is organized through null-space task prioritization and activation of set-based constraints near joint or obstacle limits. This is a markedly different architecture from direct joint-by-joint neural teleoperation.

NOIR formalizes this division of labor as hierarchical shared autonomy. SSVEP selects the object, MI selects the skill and parameters, and the robot executes parameterized primitives such as Pick(obj-A) or MoveTo(x, y, z) (Zhang et al., 2023, Kim et al., 25 Nov 2025). NOIR 2.0 extends this with faster and more accurate brain decoding algorithms, continuous closed-loop parameter control in 3D, and few-shot/one-shot robot learning based on VLMs and DINOv2 (Kim et al., 25 Nov 2025). The robot thus absorbs much of the low-level burden, while the user remains strategically in control.

Several recent systems push this principle further. An AR BRI for generalist robot arm manipulation combines gaze-based object selection, MI-based “Place”/“Use” action control, OWLv2 object detection, and a fine-tuned To.5 Vision-Language-Action model on a Franka Emika Panda arm. Safety is handled through Operational Space Control Barrier Functions, and the robot executes at 12.5 Hz while a joint controller runs at 1000 Hz (Zhang et al., 15 Jun 2026). A zero-touch grasping system uses MI decoding, AR neurofeedback, fiducial-based object confirmation, and vision-based pose estimation, with object pose in the robot base frame computed as

$^{B}\mathbf{T}_O = \,^{B}\mathbf{T}_E \, ^{E}\mathbf{T}_C \, ^{C}\mathbf{T}_O$

before waypoint-based autonomous grasp execution (Wang et al., 25 Sep 2025).

At the opposite end of the morphology spectrum, a soft-robot BRI uses MI to move a virtual Cartesian attractor, and a new Cartesian impedance controller drives the compliant end effector toward that attractor despite underactuation and non-affine actuation (Stölzle et al., 2024). The attractor update is discrete:

$x^{at}(k)=x^{at}(k-1)+\Delta_x\, s(k)\, e_a(k),$

where $N=14$0 selects the axis and $N=14$1 the sign (Stölzle et al., 2024). Here, safe operation depends not only on decoding but on physical compliance and controller design.

A related direction is brain-inspired cooperative shared control. One framework uses spiking neural networks for robotic arm and wheeled robot control, with perception, decision, and control modules connected through behavioral state machines (Yang et al., 2022). Rather than demanding high-bandwidth neural control, it assumes that noisy, low-information neural signals specify coarse intent and that adaptive SNN controllers handle the fine motion details.

4. Robotic embodiments and application domains

BRI has been demonstrated on a wide range of platforms. Social robots include Rapiro for gesture replay (Abiri et al., 2017) and Nao6 for exercise mimicry in rehabilitation sessions (Bettosi et al., 14 Sep 2025). Humanoid mobile robots include NAO in real-time natural-scene SSVEP navigation (Aznan et al., 2018) and Pepper in telepresence scenarios (Beraldo et al., 2017). Manipulation platforms include the Kinova Jaco2 assistive arm (Arrichiello et al., 2019), the Franka Emika Panda and PAL Tiago in NOIR (Zhang et al., 2023), a planar architected soft robot with four Handed Shearing Auxetic rods (Stölzle et al., 2024), a 6-DOF Arduino-controlled arm (Satam et al., 2024), and robot swarms of virtual or physical GRITSBot robots (Canal et al., 2022).

Application domains follow the categories identified in the review: assistance, healthcare, human-centric technology, entertainment, safety, military, social interaction, education, and mental recovery/meditation (Zhang et al., 2024). Rehabilitation and assistance are especially prominent. The social-robot gesture-control study explicitly frames the platform as a human-robot-interaction framework that may be combined with neurofeedback mechanisms to enhance the cognitive capability of persons with dementia (Abiri et al., 2017). The exercise-mimicry system was evaluated with 14 participants, including physiotherapists and hemiparetic patients recovering from stroke or other injuries, and was motivated by rapport-building during rehabilitation exercises (Bettosi et al., 14 Sep 2025). Assistive manipulation papers target users with severe motion disabilities (Arrichiello et al., 2019), while domestic-service systems such as DeepBrain focus on elderly users who can mildly command a robot to perform basic household tasks (Wu et al., 2020).

Another visible trend is the migration from single-user, single-robot settings to more general interaction ecologies. BRIEDGE is explicitly designed for multi-brain to multi-robot interaction, where multiple users with heterogeneous EEG devices send semantic commands through an edge server to multiple intelligent agents in parallel (Ouyang et al., 2024). NOIR targets 20 household activities spanning cooking, cleaning, personal care, and entertainment (Zhang et al., 2023). The AR BRI for generalist manipulation evaluates multi-step activities of daily living such as drinking, using a drawer, and operating an oven (Zhang et al., 15 Jun 2026). The cumulative pattern is a shift from proof-of-concept actuation toward longer-horizon, context-sensitive task execution.

5. Evaluation, reported performance, and user studies

The review notes that more than 60% of studies use performance metrics, while UX metrics, signal analyses, surveys, and interviews are also common (Zhang et al., 2024). Typical measures include classification accuracy, information transfer rate (ITR), task completion, command count, time, workload, and subjective usability. For SSVEP systems, ITR is often computed as

$N=14$2

with $N=14$3 the number of choices, $N=14$4 the accuracy, and $N=14$5 the decision time in minutes (Aznan et al., 2018).

System	Task	Reported result
Social robot gesture control	1-D cursor control via imagined finger motion	100% horizontal, 83.3% ± 11.7% vertical (Abiri et al., 2017)
Variable natural-scene SSVEP navigation	Real-time humanoid robot navigation	Mean online accuracy .85; ITR up to ~16 bpm (Aznan et al., 2018)
DeepBrain	4-class household-command EEG classification	0.970 with noise, 0.975 without noise (Wu et al., 2020)
Telepresence with ROS shared control	Mentally driving Pepper to targets	3.0 ± 1.3 commands per target; manual required 19.1% more commands (Beraldo et al., 2017)
Swarm control via posterior matching	Virtual and real swarm configuration selection	75.7% end-to-end accuracy; random guessing 1.67% (Canal et al., 2022)
Soft robot MI control	Setpoint regulation	66% of steps reached proximity of the setpoint; 21.5 s average response time for successful steps (Stölzle et al., 2024)
NOIR	20 everyday activities	Object 81.2%, skill 42.2%, parameter 73.9%, confirmation 100%; all users completed all tasks (Zhang et al., 2023)
NOIR 2.0	Everyday-task BRI with robot learning	Task completion time reduced 46%; human time reduced 65% (Kim et al., 25 Nov 2025)
Exercise mimicry	Robot-led rehabilitation session	12/14 sessions successful; user-rated mimicry accuracy mean 5.1, SD 1.26 (Bettosi et al., 14 Sep 2025)
AR generalist manipulation	Drinking, drawer, oven tasks	Online MI accuracy 0.86 ± 0.23; SUS 76.94; drawer and oven 100% (Zhang et al., 15 Jun 2026)
EEG-driven AR zero-touch grasping	Closed-loop grasping	MI accuracy 93.1%; AR neurofeedback ITR 21.3 bit/min; grasp success 97.2% (Wang et al., 25 Sep 2025)

These results show that reported performance depends strongly on the decoding paradigm and the level of autonomy delegated to the robot. Systems built around binary or low-cardinality commands and strong robot-side autonomy can achieve robust task completion despite modest raw MI accuracies. NOIR is illustrative: task-time skill selection is reported at 42.2%, yet the full system still completes all 20 tasks through modular decomposition, confirmation channels, and adaptive robot learning (Zhang et al., 2023). Conversely, systems that expose fine continuous control directly to the user, such as soft-robot setpoint regulation, report slower response and partial success even with carefully designed controllers (Stölzle et al., 2024).

Subjective measures are increasingly prominent. In the AR BRI for generalist manipulation, the mean SUS is 76.94, interpreted as “Good,” with particularly high ratings for “Quick to Learn,” “Easy to Learn,” and “Not Complex” (Zhang et al., 15 Jun 2026). In the exercise-mimicry study, trust and acceptance remained high and were not reduced by the introduction of BRI technology (Bettosi et al., 14 Sep 2025). In zero-touch grasping, users reported strong control and a sense of agency of 6.0/7, while AR neurofeedback improved sustained control with SCI = 0.210 (Wang et al., 25 Sep 2025). Such results indicate that BRI evaluation increasingly extends beyond classifier accuracy to include workload, rapport, engagement, and perceived agency.

6. Challenges, misconceptions, and future directions

The major challenges identified in the recent review are low spatial resolution, susceptibility to noise and artifacts, individual variability, external interference, real-time processing trade-offs, limited information bandwidth, safety and ethical concerns, comfort, cost, and the difficulty of long-term clinical validation (Zhang et al., 2024). These concerns recur across system papers. The exercise-mimicry study excluded seven participants because of headset fit or signal issues (Bettosi et al., 14 Sep 2025). The telepresence study emphasizes recovery behaviors and conservative navigation maps for safety (Beraldo et al., 2017). Zero-touch grasping addresses false activations through individualized calibration and AR-mediated feedback stabilization (Wang et al., 25 Sep 2025).

A common misconception is that BRI is equivalent to direct low-level robot teleoperation. The literature instead shows that many successful systems minimize neural bandwidth demands by combining BCI with autonomy, perception, or other modalities. Examples include ROS-based navigation with default forward motion (Beraldo et al., 2017), P300 command menus paired with inverse kinematic constraint handling (Arrichiello et al., 2019), object/skill/parameter decomposition in NOIR (Zhang et al., 2023), semantic edge communication in BRIEDGE (Ouyang et al., 2024), and gaze-plus-MI interaction in AR manipulation (Zhang et al., 15 Jun 2026). A plausible implication is that progress in BRI often comes less from maximizing raw EEG decoding alone than from redistributing control intelligently across the human, the decoder, and the robot.

Another misconception is that practical BRI requires one fixed EEG configuration. In the cited systems, EEG setups range from 128-channel saline nets in NOIR (Zhang et al., 2023), to 22 channels in AR manipulation (Zhang et al., 15 Jun 2026), to 14-channel consumer headsets (Abiri et al., 2017, Wang et al., 25 Sep 2025), to five-channel semi-dry systems in exercise mimicry (Bettosi et al., 14 Sep 2025), to only three bipolar channels in soft-robot control (Stölzle et al., 2024), and even two dry electrodes in DeepBrain (Wu et al., 2020). This suggests that channel count, paradigm selection, and autonomy design are tightly coupled engineering choices rather than a single optimization axis.

Future directions are explicit in multiple works: migration from offline playback to real-time closed-loop neurofeedback and richer 2-D/3-D gesture spaces (Abiri et al., 2017); larger subject pools, richer scene understanding, and hybrid BCIs for naturalistic navigation (Aznan et al., 2018); multi-session training and better-fitting headsets in rehabilitation contexts (Bettosi et al., 14 Sep 2025); semantic communication and compressed edge deployment for multi-user multi-robot settings (Ouyang et al., 2024); and evaluation with intended assistive populations for AR-based generalist manipulation (Zhang et al., 15 Jun 2026). NOIR 2.0 further points toward few-shot adaptation through foundation models, reporting one-shot adaptation where earlier NOIR learning required 15 demonstrations (Kim et al., 25 Nov 2025).

A separate, non-mainstream theoretical strand argues that classical BCI is fundamentally limited and proposes a “Quantum Cyborg” in which a human mind controls an artificial quantum computer through a quantum metalanguage grounded in dissipative quantum field theory (Pessa et al., 2009). This position differs sharply from the overwhelmingly dominant engineering literature, which is centered on EEG acquisition, statistical or deep decoding, and robot-side control architecture (Zhang et al., 2024). Its significance lies primarily in conceptual debate about the nature of intention, not in the design of contemporary BRI systems.

Taken together, the literature portrays BRI as a technically heterogeneous but increasingly coherent domain: EEG decoding remains central, yet high-performing systems now depend just as much on shared autonomy, multimodal feedback, robot learning, perception, and safety-aware control as on the neural decoder itself (Zhang et al., 2024, Kim et al., 25 Nov 2025).