NEURO/ADAPT

Embodied AI × Neurotechnology

Neuroadaptive Oversight for Embodied Agents

Implicit human neural feedback for safer autonomy in simulated urban environments.

Yin Kang et al.

Korea University

Manuscript under reviewContact Authors

Abstract & Motivation

Human oversight must keep pace with embodied autonomy.

Static alerts cannot adapt to changing workload or individual perception. Our system uses implicit neural responses to agent errors to personalize visual, auditory, and tactile feedback while a human supervises autonomous agents.

Embodied Tasks

Two supervision tasks in a simulated city

Simulated city interface for autonomous vehicle navigation01

Autonomous Navigation

Supervise a robotaxi through passenger pickup, routing, and traffic events.

Urban multi-agent monitoring environment02

Multi-agent Monitoring

Detect anomalies across vehicles and pedestrians under higher workload.

Architecture

Neuroadaptive closed loop

Decoded error-related potentials guide Bayesian optimization of personalized multimodal cues.

VisualAuditoryTactile
  1. 01Agent Error
  2. 02Decode ErrP
  3. 03Bayesian Optimization
  4. 04Adapt Multimodal Cues

Experimental Setup

A controlled human–embodied AI study

participants
0112
real-time EEG
0264-ch
baseline · static · adaptive modes
033

Selected Findings

Adaptive supervision improves human error perception

DetectionMore accurate

Improved error detection, especially under demanding conditions.

ResponseFaster

Quicker interventions than with static multimodal feedback.

NeuralEnhanced ErrP

Neural evidence consistent with stronger error perception.

Behavioral outcomes across workload and multimodal feedback conditions.
BehavioralBehavioral outcomes across workload and multimodal feedback conditions.
ErrP amplitude and latency across system, workload, and stimulus conditions.
NeuralErrP amplitude and latency across system, workload, and stimulus conditions.
ERP waveforms, scalp topographies, and band-limited neural dynamics.
ERP DynamicsERP waveforms, scalp topographies, and band-limited neural dynamics.
Personalized multimodal parameter optimization and convergence behavior.
OptimizationPersonalized multimodal parameter optimization and convergence behavior.
Model predictions and multimodal feature alignment across supervision conditions.
ModelModel predictions and multimodal feature alignment across supervision conditions.

Implications

A neural safety layer for human–embodied AI collaboration

The framework supports adaptive human oversight in autonomous driving, robotics, and high-load monitoring.