Pith. sign in

REVIEW 3 major objections 6 minor 93 references

Offline residual reinforcement learning lifts continuous 3D EEG motor-imagery decoding by correcting kinematic residuals without extra neural input.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-14 04:56 UTC pith:XUHG4AGY

load-bearing objection Solid offline residual-SAC post-decoder that lifts continuous 3D EEG MI kinematics ~20–40% without extra EEG; gains are real on held-out sessions but rest on target-aware training and time-indexed observations that have not been closed-loop tested. the 3 major comments →

arxiv 2607.11530 v1 pith:XUHG4AGY submitted 2026-07-13 cs.AI cs.LG

Learning Residual Kinematic Corrections for Continuous Neural Decoding via Reinforcement Learning

Jiamian Li , Niall McShane , Attila Korik , Naomi du Bois , Karl McCreadie , Leen Jabban , Benjamin Metcalfe , \"Ozg\"ur \c{S}im\c{s}ek
show 1 more author
Damien Coyle
This is my paper
classification cs.AI cs.LG
keywords brain-computer interfacemotor imageryEEGcontinuous kinematic decodingresidual reinforcement learningSoft Actor-CriticCNN-LSTMvirtual reality
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Continuous 3D motor imagery from non-invasive EEG is still limited by residual trajectory errors even after strong CNN–LSTM decoders. This paper shows that a second-stage offline residual reinforcement-learning agent, trained only on the decoder’s predicted kinematics and never on raw EEG, can systematically correct those residuals. Across ten participants and both 2D and immersive VR feedback, the hybrid system raises mean trajectory correlation by 41.5% and 21.2% while cutting RMSE by roughly 40%, and the corrected policy remains usable when frozen and applied to later sessions. Because the correction layer never touches the neural pipeline and needs no online user interaction during training, the approach offers a practical way to push existing continuous BCI decoders past their current performance ceiling for rehabilitation, prosthetics and virtual control.

Core claim

A frozen CNN–LSTM continuous 3D motor-imagery decoder still leaves systematic kinematic residuals; an offline residual Soft-Actor-Critic agent that observes only the decoder’s velocity outputs, prior corrected velocity and time, and is rewarded solely by Euclidean distance to the target trajectory, raises mean Pearson correlation from 0.5076 to 0.7181 (2D) and 0.6420 to 0.7780 (VR) while cutting RMSE by approximately 40%, without any additional EEG features or online policy updates.

What carries the argument

Two-stage residual RL architecture: a calibrated CNN–LSTM produces base velocities; an offline SAC agent outputs a dimension-scaled residual velocity that is added to the base prediction, with reward equal to negative Euclidean distance to the hidden target trajectory and observations deliberately free of EEG.

Load-bearing premise

Residual corrections learned offline from the first session’s decoder outputs and targets stay valid when the agent is frozen and applied to later sessions, even though the agent relies on an explicit time stamp and carefully tuned per-axis action scales.

What would settle it

Deploy the identical frozen residual agent in closed-loop real-time 3D motor-imagery control with participants who receive only the corrected trajectory as feedback; if the correlation and RMSE gains collapse relative to the offline numbers, or if removing the time observation erases the advantage, the central claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper proposes a two-stage offline framework (CNN–LSTM–RL / FRL) for continuous 3D motor-imagery decoding from EEG: a frozen CNN–LSTM first predicts limb velocities from ERSP features, then a Soft Actor-Critic agent applies residual kinematic corrections using only decoder outputs, previous corrected velocity, and normalized time—without EEG input to the RL agent. The agent is trained on Session 1 with reward equal to negative Euclidean distance to target velocities and is frozen for evaluation on later sessions under 2D and VR feedback. On ten participants, FRL raises mean Pearson correlation versus fixed CNN–LSTM (FDG) from 0.5076 to 0.7181 (2D) and 0.6420 to 0.7780 (VR), and reduces RMSE by ~38–40%, with planned pairwise tests and Holm–Bonferroni correction. The authors argue this residual offline correction improves continuous 3D BCI decoding without extra neural data and is relevant to neurorehabilitation, prosthetics, and virtual interaction.

Significance. If the offline residual-correction gains hold under fair baselines and transfer appropriately, the work offers a practical, modular way to improve continuous kinematic BCI outputs without retraining the neural decoder or feeding noisy EEG into the corrector—an attractive engineering pattern for longitudinal non-invasive BCIs. Strengths include participant-level averaging, separate 2D/VR analyses, omnibus ANOVA/Friedman tests, planned pairwise comparisons with multiplicity control, and normality/symmetry diagnostics (Appendix D). The residual-RL framing and decoupling of correction from EEG are useful contributions relative to prior CNN–LSTM 3D MI work. Significance is currently limited by the cued four-target offline open-loop setting, dependence on explicit trial time, and the lack of a simple supervised residual baseline that would show whether SAC is necessary for the reported gains.

major comments (3)
  1. [Abstract; Secs. 2.2–2.4, 3.1, 5.2; Tables 1–2] The central performance claim (FRL vs FDG/SAT/WSR in Tables 1–2 and Figs. 2–3) is an offline open-loop reconstruction result. SAC is trained with r_t = −d(v_t, v*_t) that uses true target velocities (Sec. 2.2), while observations include explicit normalized time t (Secs. 2.2, 2.4). Section 5.2 already notes that removing t collapses performance and that residual action scales are critical. The manuscript never closes the loop: residual actions do not drive feedback or alter subsequent EEG. Application-facing claims in the Abstract and Conclusions (neurorehabilitation, prosthetics, virtual interaction; “scalable framework enhances 3D BCI MI decoding”) should be scoped strictly to offline residual correction of cued home-to-target segments, or supported by closed-loop evidence. At minimum, quantify how much of the gain depends on t and on Optuna-tuned dimension-specific scales (App. C).
  2. [Secs. 2.2–2.4, 3.1; Tables 1–2] There is no comparison to a supervised residual corrector trained on the same inputs (v^BCI_t, v_{t−1}, t) with the same objectives (Euclidean velocity error and/or correlation). Because the reward is target-aware and Optuna maximizes Pearson correlation (Sec. 3.1), the reported FRL gains may be achievable by residual regression/MLP without SAC, entropy regularization, or RL-specific machinery. Without this baseline, the claim that residual reinforcement learning is the operative mechanism—rather than offline residual fitting of systematic decoder bias—is not established. A supervised residual control (and, if feasible, a residual policy trained without access to v* at reward time) is load-bearing for the paper’s methodological framing.
  3. [Sec. 3.1; Appendix C; Tables 1–2] Hyperparameter search (200 Optuna trials) explicitly maximizes the same correlation coefficient later used as a primary evaluation metric (Sec. 3.1, App. C), and residual scales are free parameters per axis. Cross-session FRL (train Session 1, test S2–S5) mitigates pure within-session overfitting, but the search still selects policies for the reported metric family. Please report: (i) whether Optuna used only the Session-1 CV folds and never test sessions; (ii) sensitivity of FRL gains to fixed vs tuned scales; and (iii) correlation/RMSE when scales and SAC hyperparameters are chosen by RMSE alone or by a held-out criterion. This is needed to show that the large relative gains (+41.5% / +21.2% r; ~40% RMSE) are not an artifact of metric-aligned search.
minor comments (6)
  1. [Introduction, contributions (1)] Typo: “CNN–LTSM” should be “CNN–LSTM” in the contributions paragraph of the Introduction.
  2. [Table 2] Table 2 VR SAT X-axis SD is reported as 0.6610, which is implausible relative to other SDs (~0.06); likely a decimal typo (0.0661?). Please verify all table entries.
  3. [Sec. 2.3, Eqs. (1)–(3)] Equation (1) has a misplaced bracket/comma in the entropy term; Eq. (2)–(3) similarly have punctuation that obscures the soft Bellman form. Clean the SAC equations for readability.
  4. [Sec. 2.4; Figure 1] Figure 1 is referenced as the two-stage architecture but the manuscript text does not fully specify online blending (“trajectory assistance blends decoded BCI trajectory with the target trajectory”) quantitatively; state the blend weights if used during data collection.
  5. [Secs. 2.1, 3.2] Clarify whether velocities or positions are correlated/RMSE’d in Eqs. (7)–(8); the text mixes “trajectory,” “points,” and “velocities.” Align notation with the decoder output (velocities) and any integration used for plotting.
  6. [Appendix B] Appendix B is titled “Reach Targets” but contains no target coordinates or figure in the provided text; ensure the appendix content is complete.

Circularity Check

2 steps flagged

Reward and Optuna both optimize the same correlation/RMSE quantities later reported as gains; cross-session FRL still supplies non-tautological held-out evidence.

specific steps
  1. fitted input called prediction [Section 3.1 (Hyperparameters Optimization); also Abstract / Tables 1–2]
    "We use Optuna[Akiba et al., 2019] for RL hyperparameters optimization with 200 searching trials, aiming to maximize the correlation coefficient between corrected and target trajectories. ... The tuned parameters, which include ... scale of residual actions in three dimensions ..."

    Hyperparameters—including dimension-specific residual action scales—are chosen by maximizing Pearson correlation on Session-1 cross-validation. The paper’s headline result is the same Pearson correlation (and related RMSE) of the corrected trajectories. The reported r gains are therefore partly the quantity the search was instructed to maximize, not a fully independent out-of-objective prediction.

  2. self definitional [Section 2.2 (reward); Section 3.2 (RMSE definition)]
    "The reward function R(st, at) is designed to minimize the error between corrected and target trajectory. It is defined as: rt =−d(vt, v∗t), where d(·,·) is a Euclidean distance metric between the corrected velocities vt and the target velocities v∗t."

    Training reward is negative Euclidean distance to the true target velocity trajectory. Evaluation RMSE is the root-mean-square of the same pointwise velocity errors. Minimizing the reward therefore optimizes the RMSE metric by construction; reported RMSE reductions (e.g., 0.0890→0.0532) restate the training objective rather than an independent derived quantity. Cross-session freeze of the agent limits but does not remove this alignment.

full rationale

This is an empirical residual-correction paper, not a first-principles derivation. The load-bearing performance claim is FRL (Session-1 SAC frozen) vs FDG/SAT/WSR on later sessions, quantified by Pearson r and RMSE. Two mild circularities exist by construction: (1) the SAC reward is negative Euclidean distance to the true target velocities—the same error family that RMSE evaluates—so RMSE reductions are the training objective, not an independent prediction; (2) Optuna selects residual scales and SAC hyperparameters by maximizing the same Pearson correlation later reported as the primary success metric. These are standard train-to-metric practices, not self-definitional uniqueness theorems or self-citation chains. The central generalization result is not forced: the residual policy is fit only on Session 1 (with targets used solely for reward, not as observations) and evaluated frozen on Sessions 2–5 under FDG-style deployment, so failure to transfer was possible. Self-citations to McShane et al. supply the baseline CNN–LSTM, not a uniqueness premise. Score 3 reflects metric-aligned fitting without collapse of the held-out claim.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 1 invented entities

The central claim rests on a standard residual-control POMDP, SAC, and a frozen CNN–LSTM baseline, plus several paper-specific modeling choices: offline rewards from known targets, no EEG in the RL observation, explicit time in the state, and heavily tuned per-axis residual scales. Free parameters are numerous because Optuna jointly fits SAC and residual scales to maximize correlation on session-1 data. Invented entities are methodological constructs (two-stage CNN–LSTM–RL / FRL), not new physical objects.

free parameters (6)
  • Residual action scales (x, y, z)
    Dimension-specific scale factors on tanh-squashed SAC actions; searched Uniform[0.1,1.0] per axis and critical to performance (authors note small scales under-correct, large scales destabilize).
  • SAC actor/critic learning rates
    Log-uniform [1e-5, 1e-3]; fitted via Optuna on correlation.
  • Hidden layer dimension, actor/critic depth, batch size
    Categorical/integer architecture choices {128,256,512}, layers [2,4], batch {64,128,256} selected by 200-trial Optuna search.
  • Discount γ and SAC target smoothing coefficient
    Uniform [0.90,0.999] and [0.005,0.05]; free RL hyperparameters tuned to the same correlation objective.
  • Trial-specific min-max normalization range [0, 0.3]
    Kinematic normalization bounds chosen for the dataset; affects residual magnitudes and reward scale.
  • CNN–LSTM architecture and ERSP window (40×40, 0–40 Hz, 16 ms steps, 17 channels)
    Baseline decoder design choices inherited/extended from prior work; residual gains are defined relative to this fixed baseline.
axioms (5)
  • domain assumption Residual kinematic errors of a frozen CNN–LSTM decoder are systematic enough that an offline policy conditioned only on v_BCI, v_{t-1}, and t can improve later sessions.
    Core premise of the two-stage framework (Abstract, §1, §2.4); if residuals are mostly non-stationary noise, FRL cannot generalize.
  • domain assumption Target velocities v* are available offline for reward computation but unavailable to the agent at observation time (POMDP residual control).
    Section 2.2 defines r_t = −d(v_t, v*_t) while o_t = [v_BCI_t, v_{t−1}, t].
  • standard math Soft Actor-Critic with entropy regularization and reparameterized tanh policies is an appropriate optimizer for continuous residual velocities.
    Section 2.3 imports SAC (Haarnoja et al.) with a scaled log-probability correction.
  • ad hoc to paper Explicit normalized time t is a valid observation for cued home-to-target trials and is necessary for strong residual performance.
    Authors state performance declines significantly without t (§5.2); this is paradigm-specific and limits asynchronous BCI.
  • domain assumption Pearson correlation and RMSE on velocity trajectories are adequate primary measures of continuous MI decoding quality for the claimed applications.
    Sections 3.2–4; no online task success or user workload metrics are primary endpoints.
invented entities (1)
  • CNN–LSTM–RL two-stage residual kinematic corrector (FRL) no independent evidence
    purpose: Name the offline residual SAC agent that corrects frozen CNN–LSTM velocity outputs without EEG input.
    Methodological construct combining known residual RL with a specific BCI decoder pipeline; independent evidence is the reported multi-session metrics, not an external physical prediction.

pith-pipeline@v1.1.0-grok45 · 19704 in / 3857 out tokens · 50556 ms · 2026-07-14T04:56:47.925540+00:00 · methodology

0 comments
read the original abstract

Decoding continuous three-dimensional (3D) motor imagery (MI) using non-invasive electroencephalography (EEG)-based brain--computer interfaces (BCIs) remains challenging due to signal variability and residual decoding errors. Deep learning architectures such as convolutional neural network--long short-term memory (CNN--LSTM) models can capture spatial and temporal dynamics for continuous kinematic decoding; however, systematic residual errors persist in predicted trajectories. We propose a two-stage decoding framework that applies reinforcement learning (RL) to perform residual kinematic correction on the outputs of a CNN--LSTM decoder (CNN--LSTM--RL). The RL agent is trained offline without direct EEG input and instead operates on predicted kinematic trajectories to optimize movement accuracy relative to target trajectories. Decoding performance was quantified using Pearson correlation coefficients ($r$) and Root Mean Square Errors (RMSE) along the $x, y$, and $z$ axes. Compared to CNN--LSTM applied alone, CNN--LSTM--RL improved the mean correlation from $0.5076$ to $0.7181$ ($p = 0.0005$) in 2D and from $0.6420$ to $0.7780$ ($p = 0.0059$) in VR, with relative gains of $41.5\%$ and $21.2\%$, respectively. Correspondingly, RMSE was reduced from $0.0890$ to $0.0532$ (2D, $p < 0.0001$) and from $0.0714$ to $0.0441$ (VR, $p < 0.0001$), representing relative reductions of $40.2\%$ and $38.2\%$. These findings demonstrate that this scalable framework enhances 3D BCI MI decoding by correcting kinematic errors via offline residual RL without extra neural data, advancing neurorehabilitation, prosthetics, and virtual interaction.

Figures

Figures reproduced from arXiv: 2607.11530 by Attila Korik, Benjamin Metcalfe, Damien Coyle, Jiamian Li, Karl McCreadie, Leen Jabban, Naomi Du Bois, Niall McShane, \"Ozg\"ur \c{S}im\c{s}ek.

Figure 1
Figure 1. Figure 1: The architecture of our proposed two-stage framework. [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Pearson correlation coefficient of FRL comparison with three training strategies of BCI [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Root mean square errors comparison of FRL with three training strategies of BCI decoder [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

93 extracted references · 4 linked inside Pith

  1. [3]

    Proceedings of the 25th

    Optuna: A Next-generation Hyperparameter Optimization Framework , author=. Proceedings of the 25th

  2. [4]

    Frontiers in neuroscience , volume=

    Combining brain--computer interfaces and assistive technologies: state-of-the-art and challenges , author=. Frontiers in neuroscience , volume=. 2010 , publisher=

    2010

  3. [5]

    Annals of clinical and translational neurology , volume=

    Brain-computer interfaces for post-stroke motor rehabilitation: a meta-analysis , author=. Annals of clinical and translational neurology , volume=. 2018 , publisher=

    2018

  4. [6]

    Nature , volume=

    High-performance brain-to-text communication via handwriting , author=. Nature , volume=. 2021 , publisher=

    2021

  5. [7]

    Nature , volume=

    A high-performance speech neuroprosthesis , author=. Nature , volume=. 2023 , publisher=

    2023

  6. [8]

    Nature , volume=

    Speech synthesis from neural decoding of spoken sentences , author=. Nature , volume=. 2019 , publisher=

    2019

  7. [9]

    IEEE Transactions on Computational Intelligence and AI in Games , volume=

    Games, gameplay, and BCI: the state of the art , author=. IEEE Transactions on Computational Intelligence and AI in Games , volume=. 2013 , publisher=

    2013

  8. [10]

    IEEE Transactions on Computational Intelligence and AI in Games , volume=

    Experiencing BCI control in a popular computer game , author=. IEEE Transactions on Computational Intelligence and AI in Games , volume=. 2013 , publisher=

    2013

  9. [11]

    Nature , volume=

    Reach and grasp by people with tetraplegia using a neurally controlled robotic arm , author=. Nature , volume=. 2012 , publisher=

    2012

  10. [12]

    Science robotics , volume=

    Noninvasive neuroimaging enhances continuous neural tracking for robotic device control , author=. Science robotics , volume=. 2019 , publisher=

    2019

  11. [13]

    The Lancet , volume=

    Restoration of reaching and grasping movements through brain-controlled muscle stimulation in a person with tetraplegia: a proof-of-concept demonstration , author=. The Lancet , volume=. 2017 , publisher=

    2017

  12. [14]

    TRENDS in Neurosciences , volume=

    Brain--machine interfaces: past, present and future , author=. TRENDS in Neurosciences , volume=. 2006 , publisher=

    2006

  13. [15]

    Journal of neural engineering , volume=

    A review of classification algorithms for EEG-based brain--computer interfaces: a 10 year update , author=. Journal of neural engineering , volume=. 2018 , publisher=

    2018

  14. [16]

    Journal of neural engineering , volume=

    A comprehensive review of EEG-based brain--computer interface paradigms , author=. Journal of neural engineering , volume=. 2019 , publisher=

    2019

  15. [17]

    2022 IEEE International Conference on Metrology for Extended Reality, Artificial Intelligence and Neural Engineering (Metroxraine) , pages=

    Online 3D motion decoder BCI for embodied virtual reality upper limb control: A pilot study , author=. 2022 IEEE International Conference on Metrology for Extended Reality, Artificial Intelligence and Neural Engineering (Metroxraine) , pages=. 2022 , organization=

    2022

  16. [18]

    2023 IEEE International Conference on Metrology for eXtended Reality, Artificial Intelligence and Neural Engineering (MetroXRAINE) , pages=

    Decoding motion trajectories in an upper limb BCI: Linear regression vs deep learning , author=. 2023 IEEE International Conference on Metrology for eXtended Reality, Artificial Intelligence and Neural Engineering (MetroXRAINE) , pages=. 2023 , organization=

    2023

  17. [19]

    Frontiers in neuroscience , volume=

    A comprehensive review on motion trajectory reconstruction for EEG-based brain-computer interface , author=. Frontiers in neuroscience , volume=. 2023 , publisher=

    2023

  18. [20]

    Journal of neuroscience , volume=

    Reconstructing three-dimensional hand movements from noninvasive electroencephalographic signals , author=. Journal of neuroscience , volume=. 2010 , publisher=

    2010

  19. [21]

    IEEE Transactions on Neural Systems and Rehabilitation Engineering , year=

    Decoding the variable velocity of lower-limb stepping movements from EEG , author=. IEEE Transactions on Neural Systems and Rehabilitation Engineering , year=

  20. [22]

    IEEE Journal of Biomedical and Health Informatics , volume=

    Continuous bimanual trajectory decoding of coordinated movement from EEG signals , author=. IEEE Journal of Biomedical and Health Informatics , volume=. 2022 , publisher=

    2022

  21. [23]

    IEEE Transactions on Neural Systems and Rehabilitation Engineering , volume=

    Decoding hand movement types and kinematic information from electroencephalogram , author=. IEEE Transactions on Neural Systems and Rehabilitation Engineering , volume=. 2021 , publisher=

    2021

  22. [24]

    1998 , publisher=

    Reinforcement learning: An introduction , author=. 1998 , publisher=

    1998

  23. [25]

    The International Journal of Robotics Research , volume=

    Reinforcement learning in robotics: A survey , author=. The International Journal of Robotics Research , volume=. 2013 , publisher=

    2013

  24. [26]

    Artificial intelligence review , volume=

    Reinforcement learning in robotic applications: a comprehensive survey , author=. Artificial intelligence review , volume=. 2022 , publisher=

    2022

  25. [27]

    Journal of Neural Engineering , volume=

    Error-related potential-based shared autonomy via deep recurrent reinforcement learning , author=. Journal of Neural Engineering , volume=. 2022 , publisher=

    2022

  26. [28]

    Scientific reports , volume=

    Intrinsic interactive reinforcement learning--using error-related potentials for real world human-robot interaction , author=. Scientific reports , volume=. 2017 , publisher=

    2017

  27. [29]

    2019 international conference on robotics and automation (ICRA) , pages=

    Residual reinforcement learning for robot control , author=. 2019 international conference on robotics and automation (ICRA) , pages=. 2019 , organization=

    2019

  28. [31]

    IEEE Robotics and Automation Letters , volume=

    Residual learning from demonstration: Adapting dmps for contact-rich manipulation , author=. IEEE Robotics and Automation Letters , volume=. 2022 , publisher=

    2022

  29. [32]

    Breakthroughs in statistics: Methodology and distribution , pages=

    Individual comparisons by ranking methods , author=. Breakthroughs in statistics: Methodology and distribution , pages=. 1992 , publisher=

    1992

  30. [33]

    2013 , publisher=

    Statistical power analysis for the behavioral sciences , author=. 2013 , publisher=

    2013

  31. [34]

    Frontiers in Human Neuroscience , volume=

    Hybrid brain-computer interface using error-related potential and reinforcement learning , author=. Frontiers in Human Neuroscience , volume=. 2025 , publisher=

    2025

  32. [35]

    Expert Systems with Applications , volume=

    Brain-controlled operator model-driven deep reinforcement learning for adaptive brain-machine collaborative control , author=. Expert Systems with Applications , volume=. 2026 , publisher=

    2026

  33. [36]

    Applied Sciences , volume=

    EEG-based inverse reinforcement learning for safety-oriented global path planning in dynamic environments , author=. Applied Sciences , volume=. 2025 , publisher=

    2025

  34. [37]

    Neurocomputing , volume=

    Accelerating reinforcement learning using EEG-based implicit human feedback , author=. Neurocomputing , volume=. 2021 , publisher=

    2021

  35. [38]

    Proceedings of the 6th ACM Workshop on Wearable Systems and Applications , pages=

    Human-in-the-loop rl with an eeg wearable headset: On effective use of brainwaves to accelerate learning , author=. Proceedings of the 6th ACM Workshop on Wearable Systems and Applications , pages=

  36. [39]

    Annual Review of Control, Robotics, and Autonomous Systems , volume=

    Deep reinforcement learning for robotics: A survey of real-world successes , author=. Annual Review of Control, Robotics, and Autonomous Systems , volume=. 2025 , publisher=

    2025

  37. [40]

    IEEE Transactions on Systems, Man, and Cybernetics: Systems , year=

    Reinforcement learning methods for assistive and rehabilitation robotic systems: A survey , author=. IEEE Transactions on Systems, Man, and Cybernetics: Systems , year=

  38. [41]

    Computers in Biology and Medicine , volume=

    Enhancing motor imagery EEG signal decoding through machine learning: A systematic review of recent progress , author=. Computers in Biology and Medicine , volume=. 2025 , publisher=

    2025

  39. [42]

    Brain Hemorrhages , year=

    Neural Decoding for EEG-BCI: From Conventional Machine Learning to Deep Learning Models , author=. Brain Hemorrhages , year=

  40. [44]

    Scandinavian journal of statistics , pages=

    A simple sequentially rejective multiple test procedure , author=. Scandinavian journal of statistics , pages=. 1979 , publisher=

    1979

  41. [45]

    Journal of neuroengineering and rehabilitation , volume=

    Applying a brain-computer interface to support motor imagery practice in people with stroke for upper limb recovery: a feasibility study , author=. Journal of neuroengineering and rehabilitation , volume=. 2010 , publisher=

    2010

  42. [46]

    Communications medicine , year=

    Advancing EEG-based assessment of consciousness and cognition in prolonged disorders of consciousness , author=. Communications medicine , year=

  43. [47]

    Electroencephalography and clinical neurophysiology , volume=

    Auditory event-related dynamics of the EEG spectrum and effects of exposure to tones , author=. Electroencephalography and clinical neurophysiology , volume=. 1993 , publisher=

    1993

  44. [48]

    1992 , publisher=

    ANOVA: Repeated measures , author=. 1992 , publisher=

    1992

  45. [49]

    Journal of the american statistical association , volume=

    The use of ranks to avoid the assumption of normality implicit in the analysis of variance , author=. Journal of the american statistical association , volume=. 1937 , publisher=

    1937

  46. [50]

    u diger Rupp, Gernot R M \

    J d R Mill \'a n, R \"u diger Rupp, Gernot R M \"u ller-Putz, Roderick Murray-Smith, Claudio Giugliemma, Michael Tangermann, Carmen Vidaurre, Febo Cincotti, Andrea K \"u bler, Robert Leeb, et al. Combining brain--computer interfaces and assistive technologies: state-of-the-art and challenges. Frontiers in neuroscience, 4: 0 161, 2010

    2010

  47. [51]

    Applying a brain-computer interface to support motor imagery practice in people with stroke for upper limb recovery: a feasibility study

    Girijesh Prasad, Pawel Herman, Damien Coyle, Suzanne McDonough, and Jacqueline Crosbie. Applying a brain-computer interface to support motor imagery practice in people with stroke for upper limb recovery: a feasibility study. Journal of neuroengineering and rehabilitation, 7 0 (1): 0 60, 2010

    2010

  48. [52]

    Brain-computer interfaces for post-stroke motor rehabilitation: a meta-analysis

    Mar \' a A Cervera, Surjo R Soekadar, Junichi Ushiba, Jos \'e del R Mill \'a n, Meigen Liu, Niels Birbaumer, and Gangadhar Garipelli. Brain-computer interfaces for post-stroke motor rehabilitation: a meta-analysis. Annals of clinical and translational neurology, 5 0 (5): 0 651--663, 2018

    2018

  49. [53]

    High-performance brain-to-text communication via handwriting

    Francis R Willett, Donald T Avansino, Leigh R Hochberg, Jaimie M Henderson, and Krishna V Shenoy. High-performance brain-to-text communication via handwriting. Nature, 593 0 (7858): 0 249--254, 2021

    2021

  50. [54]

    A high-performance speech neuroprosthesis

    Francis R Willett, Erin M Kunz, Chaofei Fan, Donald T Avansino, Guy H Wilson, Eun Young Choi, Foram Kamdar, Matthew F Glasser, Leigh R Hochberg, Shaul Druckmann, et al. A high-performance speech neuroprosthesis. Nature, 620 0 (7976): 0 1031--1036, 2023

    2023

  51. [55]

    Speech synthesis from neural decoding of spoken sentences

    Gopala K Anumanchipalli, Josh Chartier, and Edward F Chang. Speech synthesis from neural decoding of spoken sentences. Nature, 568 0 (7753): 0 493--498, 2019

    2019

  52. [56]

    Advancing eeg-based assessment of consciousness and cognition in prolonged disorders of consciousness

    Naomi Du Bois, Attila Korik, Stephanie Hodge, Leah Hudson, Ainjila S Elahi, Alain Bigirimana, Natalie Dayan, Jose M Sanchez-Bornot, Alison McCann, Kudret Yelden, et al. Advancing eeg-based assessment of consciousness and cognition in prolonged disorders of consciousness. Communications medicine, 2026

    2026

  53. [57]

    Games, gameplay, and bci: the state of the art

    David Marshall, Damien Coyle, Shane Wilson, and Michael Callaghan. Games, gameplay, and bci: the state of the art. IEEE Transactions on Computational Intelligence and AI in Games, 5 0 (2): 0 82--99, 2013

    2013

  54. [58]

    Experiencing bci control in a popular computer game

    Bram van de Laar, Hayrettin G \"u rk \"o k, Danny Plass-Oude Bos, Mannes Poel, and Anton Nijholt. Experiencing bci control in a popular computer game. IEEE Transactions on Computational Intelligence and AI in Games, 5 0 (2): 0 176--184, 2013

    2013

  55. [59]

    Reach and grasp by people with tetraplegia using a neurally controlled robotic arm

    Leigh R Hochberg, Daniel Bacher, Beata Jarosiewicz, Nicolas Y Masse, John D Simeral, Joern Vogel, Sami Haddadin, Jie Liu, Sydney S Cash, Patrick Van Der Smagt, et al. Reach and grasp by people with tetraplegia using a neurally controlled robotic arm. Nature, 485 0 (7398): 0 372--375, 2012

    2012

  56. [60]

    Restoration of reaching and grasping movements through brain-controlled muscle stimulation in a person with tetraplegia: a proof-of-concept demonstration

    A Bolu Ajiboye, Francis R Willett, Daniel R Young, William D Memberg, Brian A Murphy, Jonathan P Miller, Benjamin L Walter, Jennifer A Sweet, Harry A Hoyen, Michael W Keith, et al. Restoration of reaching and grasping movements through brain-controlled muscle stimulation in a person with tetraplegia: a proof-of-concept demonstration. The Lancet, 389 0 (10...

    2017

  57. [61]

    Noninvasive neuroimaging enhances continuous neural tracking for robotic device control

    Bradley J Edelman, Jianjun Meng, Daniel Suma, Claire Zurn, Eric Nagarajan, Bryan S Baxter, Christopher C Cline, and BJSR He. Noninvasive neuroimaging enhances continuous neural tracking for robotic device control. Science robotics, 4 0 (31): 0 eaaw6844, 2019

    2019

  58. [62]

    Brain--machine interfaces: past, present and future

    Mikhail A Lebedev and Miguel AL Nicolelis. Brain--machine interfaces: past, present and future. TRENDS in Neurosciences, 29 0 (9): 0 536--546, 2006

    2006

  59. [63]

    A review of classification algorithms for eeg-based brain--computer interfaces: a 10 year update

    Fabien Lotte, Laurent Bougrain, Andrzej Cichocki, Maureen Clerc, Marco Congedo, Alain Rakotomamonjy, and Florian Yger. A review of classification algorithms for eeg-based brain--computer interfaces: a 10 year update. Journal of neural engineering, 15 0 (3): 0 031005, 2018

    2018

  60. [64]

    A comprehensive review of eeg-based brain--computer interface paradigms

    Reza Abiri, Soheil Borhani, Eric W Sellers, Yang Jiang, and Xiaopeng Zhao. A comprehensive review of eeg-based brain--computer interface paradigms. Journal of neural engineering, 16 0 (1): 0 011001, 2019

    2019

  61. [65]

    Reconstructing three-dimensional hand movements from noninvasive electroencephalographic signals

    Trent J Bradberry, Rodolphe J Gentili, and Jos \'e L Contreras-Vidal. Reconstructing three-dimensional hand movements from noninvasive electroencephalographic signals. Journal of neuroscience, 30 0 (9): 0 3432--3437, 2010

    2010

  62. [66]

    Continuous bimanual trajectory decoding of coordinated movement from eeg signals

    Yi-Feng Chen, Ruiqi Fu, Junde Wu, Jongbin Song, Rui Ma, Yi-Chuan Jiang, and Mingming Zhang. Continuous bimanual trajectory decoding of coordinated movement from eeg signals. IEEE Journal of Biomedical and Health Informatics, 26 0 (12): 0 6012--6023, 2022

    2022

  63. [67]

    A comprehensive review on motion trajectory reconstruction for eeg-based brain-computer interface

    Pengpai Wang, Xuhao Cao, Yueying Zhou, Peiliang Gong, Muhammad Yousefnezhad, Wei Shao, and Daoqiang Zhang. A comprehensive review on motion trajectory reconstruction for eeg-based brain-computer interface. Frontiers in neuroscience, 17: 0 1086472, 2023

    2023

  64. [68]

    Decoding hand movement types and kinematic information from electroencephalogram

    Baoguo Xu, Yong Wang, Leying Deng, Changcheng Wu, Wenbing Zhang, Huijun Li, and Aiguo Song. Decoding hand movement types and kinematic information from electroencephalogram. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 29: 0 1744--1755, 2021 a

    2021

  65. [69]

    Decoding the variable velocity of lower-limb stepping movements from eeg

    Attila Korik, Naomi Du Bois, Jose Sanchez Bornot, Niall McShane, Christoph Guger, Alessandra Del Felice, Olive Lennon, and Damien Coyle. Decoding the variable velocity of lower-limb stepping movements from eeg. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 2025

    2025

  66. [70]

    Decoding motion trajectories in an upper limb bci: Linear regression vs deep learning

    Niall Mc Shane, Karl McCreadie, Darryl Charles, Attila Korik, and Damien Coyle. Decoding motion trajectories in an upper limb bci: Linear regression vs deep learning. In 2023 IEEE International Conference on Metrology for eXtended Reality, Artificial Intelligence and Neural Engineering (MetroXRAINE), pages 1039--1044. IEEE, 2023

    2023

  67. [71]

    Embodied virtual reality feedback reshapes neural representations to support continuous three-dimensional motor imagery decoding

    Niall McShane, Attila Korik, Karl McCreadie, Naomi Du Bois, Darryl Charles, and Damien Coyle. Embodied virtual reality feedback reshapes neural representations to support continuous three-dimensional motor imagery decoding. arXiv preprint arXiv:2605.29677, 2026

    Pith/arXiv arXiv 2026

  68. [72]

    Online 3d motion decoder bci for embodied virtual reality upper limb control: A pilot study

    Niall Mc Shane, Karl McCreadie, Darryl Charles, Attila Korik, and Damien Coyle. Online 3d motion decoder bci for embodied virtual reality upper limb control: A pilot study. In 2022 IEEE International Conference on Metrology for Extended Reality, Artificial Intelligence and Neural Engineering (Metroxraine), pages 697--702. IEEE, 2022

    2022

  69. [73]

    Reinforcement learning: An introduction, volume 1

    Richard S Sutton, Andrew G Barto, et al. Reinforcement learning: An introduction, volume 1. MIT press Cambridge, 1998

    1998

  70. [74]

    Reinforcement learning in robotics: A survey

    Jens Kober, J Andrew Bagnell, and Jan Peters. Reinforcement learning in robotics: A survey. The International Journal of Robotics Research, 32 0 (11): 0 1238--1274, 2013

    2013

  71. [75]

    Reinforcement learning in robotic applications: a comprehensive survey

    Bharat Singh, Rajesh Kumar, and Vinay Pratap Singh. Reinforcement learning in robotic applications: a comprehensive survey. Artificial intelligence review, 55 0 (2): 0 945--990, 2022

    2022

  72. [76]

    Deep reinforcement learning for robotics: A survey of real-world successes

    Chen Tang, Ben Abbatematteo, Jiaheng Hu, Rohan Chandra, Roberto Mart \' n-Mart \' n, and Peter Stone. Deep reinforcement learning for robotics: A survey of real-world successes. Annual Review of Control, Robotics, and Autonomous Systems, 8 0 (1): 0 153--188, 2025

    2025

  73. [77]

    Reinforcement learning methods for assistive and rehabilitation robotic systems: A survey

    Mojtaba Sharifi, Shreesh Tripathi, Yun Chen, Qiang Zhang, and Mahdi Tavakoli. Reinforcement learning methods for assistive and rehabilitation robotic systems: A survey. IEEE Transactions on Systems, Man, and Cybernetics: Systems, 2025

    2025

  74. [78]

    Intrinsic interactive reinforcement learning--using error-related potentials for real world human-robot interaction

    Su Kyoung Kim, Elsa Andrea Kirchner, Arne Stefes, and Frank Kirchner. Intrinsic interactive reinforcement learning--using error-related potentials for real world human-robot interaction. Scientific reports, 7 0 (1): 0 17562, 2017

    2017

  75. [79]

    Error-related potential-based shared autonomy via deep recurrent reinforcement learning

    Xiaofei Wang, Hsiang-Ting Chen, and Chin-Teng Lin. Error-related potential-based shared autonomy via deep recurrent reinforcement learning. Journal of Neural Engineering, 19 0 (6): 0 066023, 2022

    2022

  76. [80]

    Hybrid brain-computer interface using error-related potential and reinforcement learning

    Aline Xavier Fid \^e ncio, Felix Gr \"u n, Christian Klaes, and Ioannis Iossifidis. Hybrid brain-computer interface using error-related potential and reinforcement learning. Frontiers in Human Neuroscience, 19: 0 1569411, 2025

    2025

  77. [81]

    Brain-controlled operator model-driven deep reinforcement learning for adaptive brain-machine collaborative control

    Zichao Xu, Luzheng Bi, Zhenge Yang, Haorui Ge, Zitong Wang, Kaixuan Lian, Weijie Fei, and Peiyu Zhang. Brain-controlled operator model-driven deep reinforcement learning for adaptive brain-machine collaborative control. Expert Systems with Applications, 305: 0 130770, 2026

    2026

  78. [82]

    Human-in-the-loop rl with an eeg wearable headset: On effective use of brainwaves to accelerate learning

    Mohit Agarwal, Shyam Krishnan Venkateswaran, and Raghupathy Sivakumar. Human-in-the-loop rl with an eeg wearable headset: On effective use of brainwaves to accelerate learning. In Proceedings of the 6th ACM Workshop on Wearable Systems and Applications, pages 25--30, 2020

    2020

  79. [83]

    Accelerating reinforcement learning using eeg-based implicit human feedback

    Duo Xu, Mohit Agarwal, Ekansh Gupta, Faramarz Fekri, and Raghupathy Sivakumar. Accelerating reinforcement learning using eeg-based implicit human feedback. Neurocomputing, 460: 0 139--153, 2021 b

    2021

  80. [84]

    Eeg-based inverse reinforcement learning for safety-oriented global path planning in dynamic environments

    Hao Zhu, Jialin Wang, and Rui Gao. Eeg-based inverse reinforcement learning for safety-oriented global path planning in dynamic environments. Applied Sciences, 15 0 (11): 0 6163, 2025

    2025

Showing first 80 references.