罗骁齐

Individuals with severe motor or communication impairments may retain clear object goals but lack a reliable way to convey them to assistive robots. We present NeuralIntent, a framework that maps noninvasive electroencephalography (EEG) into continuous Neural Intent Tokens (NITs) and calibrates them to condition an embodied foundation model. NeuralIntent maps the EEG-derived representation into the policy condition space through an embodiment-specific adapter. The EEG-derived representation specifies the target object, while the embodied policy generates the actions needed to interact with it. We introduce the Embodied Intent Grounding Benchmark (EIG-Bench). On THINGS-EEG2, Qwen3-VL Layer 6 supervision improves retrieval among 200 candidates from 25.27% to 40.58%. Under strict leave-one-subject-out evaluation, calibrated NITs achieve 48.13% target selection accuracy. In a separate modular evaluation, the system achieves 65.63% task success, with all correctly selected objects subsequently grasped and lifted.

Electroencephalography (EEG) offers millisecond-level temporal resolution but limited spatial localization, whereas functional near-infrared spectroscopy (fNIRS) provides better spatial resolution but shows a hemodynamic delay of several seconds. A hybrid EEG-fNIRS brain-computer interface fuses the complementary strengths of both modalities but faces challenges from intrinsic spatiotemporal heterogeneity between electrophysiological and hemodynamic signals. To address this, this study proposes a Progressive Cross-Modal Alignment Network (PCMAN), which achieves deep EEG-fNIRS fusion and alignment through a three-layer progressive mechanism: physical alignment (a Spatiotemporal correlation Dynamic Fusion framework, SDF), semantic alignment (a Multi-Dimensional Feature Enhancement Attention network, MDFEA), and task alignment (a Cross-modal Adaptive Decoding Network, CADN). On three cognitive tasks–mental arithmetic (MA), word generation (WG), and motor imagery (MI)–PCMAN achieved accuracies of 88.18%, 81.93%, and 72.94%, outperforming state-of-the-art hybrid BCI methods.

Electroencephalography (EEG) has been widely used in emotion recognition, yet the scarcity of labeled samples remains a major obstacle to its broader application. Existing semi-supervised EEG-based emotion recognition methods mainly rely on data-driven distribution alignment strategies and have not fully explored the neurophysiological priors underlying affective neural activity. To address this issue, we propose a neuro-aware semi-supervised contrastive learning framework, termed Neuro-SSCL. It incorporates neuroscience-derived inductive biases into model design and optimization. Neurophysiological topology-aware soft contrastive learning (NT-SCL) explicitly integrates local spatial correlations induced by volume conduction and global functional connectivity topology into contrastive learning, while spectrum-adaptive guided multi-scale graph contrastive learning (SA-MGCL) enhances emotion-relevant spectral components and imposes multi-scale graph contrastive constraints. A semi-supervised multi-domain adaptation strategy further mitigates distribution shifts across labeled/unlabeled source and target domains via adversarial alignment. On SEED, SEED-IV and SEED-V under a semi-supervised cross-subject leave-one-subject-out protocol, Neuro-SSCL consistently outperforms state-of-the-art methods under different incomplete-label conditions.

Scalp EEG signals are affected by volume-conduction-induced mixing, whereas source-localized time series contain nonlinear dynamics together with redundant activity from task-irrelevant regions of interest (ROIs). This study proposes PSSF, a phase-space and state-space fusion method for EEG source localization and feature decoding. Scalp EEG is projected into cortical source space using sLORETA to obtain ROI-level source time series. Phase-space reconstruction via time-delay embedding, together with approximate, fuzzy, and permutation entropy, characterizes nonlinear source dynamics; in parallel, a state-space model performs sparse source estimation and BIC-based model selection to identify active ROIs. Features retained under active-ROI constraints are ranked by ReliefF and classified with a linear SVM. PSSF achieved mean accuracies of 77.81% on BCI Competition IV 2a (binary MI) and 98.35% on a self-collected adult ADHD dataset; ablation showed gains of 6.98 and 5.72 percentage points for fuzzy and hybrid entropy after jointly introducing phase-space reconstruction and state-space ROI screening.

Continuous affect tracking is central to closed-loop affective brain–computer interfaces and adaptive human–computer interaction. Continuous EEG–fNIRS regression is well suited to this setting because EEG captures rapid neural dynamics and fNIRS provides complementary hemodynamic information, but direct fusion often underperforms EEG-only baselines when neurovascular coupling (NVC) introduces subject- and region-dependent delays. PhysioSync addresses this by performing HRF-like temporal alignment and residual-aware reliability control before multimodal fusion. A learnable dual-gamma kernel maps EEG-derived drive to the fNIRS timeline through finite-window causal convolution, and the resulting residuals modulate the contribution of hemodynamic evidence during fusion. A linear-complexity Mamba encoder provides long-context EEG representation. On REFED, PhysioSync reduces Avg MAE by 5.8% relative to the best fusion baseline and by 7.2% relative to EEG-only when evaluated across three within-subject folds, while showing lower error in 1–3 s trajectory forecasting. With lightweight task-specific adaptations, ENTER and ZBra-music results indicate conditional applicability to discrete classification rather than uniform transferability, with performance depending on EEG coverage and HbO/HbR chromophore decomposition.