IROS 2026 Commit-to-Execute EEG–EMG–ET

NeuroCommitSSM: Decision-Centric Shared Autonomy for Safe Assistive Manipulation via EEG–EMG–ET Commit Readiness

NeuroCommitSSM predicts continuous commit readiness from synchronized EEG, EMG, and eye tracking, then converts it into stable commit events via dwell/hysteresis filtering and gates execution with feasibility-grounded HOLD–ASSIST–COMMIT (HAC) shared autonomy.

Tipu Sultan1, Param Sangani2, Kody Cool1, Pascal Sikorski2, Guangping Liu1, Hadi Akbarpour2, Madi Babaiasl1
1Saint Louis University — Aerospace & Mechanical Engineering · 2Saint Louis University — Computer Science

Overview

NeuroCommitSSM overview figure
Recognition
0.950
Action Bal-Acc (S0)
Safety
0.75
FP / 1k REST (S0)
Runtime
7.89 ms
Latency / window (A100)
System
96.9%
HIL success (Full HAC)

Metrics are reported in the paper under LOSO cross-validation and sensor-dropout scenarios (S0–S6). System-level validation uses Kinova Gen3 hardware-in-the-loop (HIL) evaluation of supervisor variants.

Abstract

Safe commit-to-execute decisions are critical for assistive robotic manipulators: the system must minimize false activations during rest while remaining responsive to user intent. Existing intent-recognition pipelines conflate window-level classification with commit readiness, yielding high false-positive rates and instability under cross-subject variability and sensor dropout. We present NeuroCommitSSM, a decision-centric framework that learns continuous commit-readiness from synchronized electroencephalography (EEG), electromyography (EMG), and eye tracking (ET) using neurophysiology-aware encoders and entropy-regularized multimodal fusion. Discrete commit events are generated via dwell/hysteresis filtering and coupled with a HOLD–ASSIST–COMMIT (HAC) supervisor that gates execution on perception and robot-state feasibility. Under leave-one-subject-out (LOSO) cross-validation (N=32, five International Classification of Functioning, Disability and Health (ICF)-aligned activities of daily living (ADL) tasks), NeuroCommitSSM achieves 0.950 action balanced accuracy and 0.75 false commits per 1000 s of rest, with graceful degradation under sensor dropout. Hardware-in-the-loop (HIL) validation on a Kinova Gen3 arm confirms safe execution under real-world constraints.

Method

Commit Readiness from EEG–EMG–ET (Model Architecture)

NeuroCommitSSM operates on synchronized tri-modal windows of length 2.0 s (T=500 samples at 250 Hz; stride 0.25 s) and propagates per-sample validity masks throughout tokenization, encoding, fusion, and pooling. For each window, the model jointly predicts (i) REST vs. ACTION, (ii) task class (T1–T5) for ACTION windows, and (iii) a continuous commit-readiness score ct∈[0,1] that is converted to stable commit events by dwell/hysteresis filtering and consumed by the HOLD–ASSIST–COMMIT supervisor.

  • Masked patch tokenization: masked patch projection (p=25 → L=20) into D=160 embeddings; patch validity suppresses unreliable tokens.
  • Encoders: EEG multiscale filterbank + virtual-channel mixing (K=6); EMG envelope/burst pathways + synergy basis (M=4); ET validity-aware salience with top-k event tokens.
  • Reliability-aware fusion: quality proxies → uncertainty weights; entropy regularization discourages single-modality dominance.
  • Auxiliary features: compact descriptors (EEG PSD + EMG time-domain + modality flags; 123-D) injected for dropout robustness.
  • Heads: action (REST/ACTION), task (T1–T5), and commit head predicting continuous readiness used by dwell/hysteresis filtering.
NeuroCommitSSM architecture figure

Pipeline

Pipeline and HAC Shared Autonomy (Data Collection → Evaluation)

The system implements an end-to-end intent-to-commit workflow that converts synchronized EEG–EMG–ET into window-level predictions and commit-to-execute decisions. Commit readiness is stabilized via dwell/hysteresis filtering and integrated into a ROS2 supervisor (HOLD–ASSIST–COMMIT) that gates execution on perception and robot-state feasibility for safe, repeatable behavior under cross-subject shift and sensor dropout.

  • Data acquisition: synchronized EEG (8ch @ 250 Hz), EMG (4ch), and eye tracking via iMotions/LSL during five ICF-aligned ADL tasks plus REST.
  • Preprocessing & windowing: deterministic filtering/repair per modality; ET validity handling; 2.0 s windows with 0.25 s stride; train-only normalization.
  • Commit formation: readiness ct → discrete commits using dwell + hysteresis (separate ON/OFF thresholds; optional cooldown) to suppress transient spikes.
  • Feasibility gating: perception feasibility fcv(t) + robot feasibility frobot(t) (IK, collisions, joint limits). Safety aborts return to HOLD.
  • Evaluation: LOSO + sensor-dropout scenarios (S0–S6) with decision-quality metrics (FP/1k REST, flaps/min, coverage) and Kinova Gen3 HIL supervisor ablations.
Tri-modal pipeline figure

Video

This page includes (i) the data-collection protocol clips for each task (T1–T5) and (ii) a Kinova Gen3 HAC shared-autonomy demonstration video. If a video does not load due to hosting/browser limitations, use the “Open” button beneath the clip to view it on GitHub.


Kinova Gen3 Demonstration (HAC Shared Autonomy)

Project Demo — HAC on Kinova Gen3

End-to-end execution under feasibility-grounded HOLD–ASSIST–COMMIT supervision with commit-to-execute decision gating.


Task Videos (Data Collection Protocol)

The following clips illustrate the data-collection protocol for each task (T1–T5), including participant interaction with the setup and the trial workflow.

T1 — Clock

Reaching toward a tabletop clock and grasping it.

T2 — Bottle

Picking up a dropped apple juice bottle and placing it back onto the table.

T3 — Fan

Pressing the fan power button to toggle it on and off.

T4 — Plant

Adjusting or repositioning a plant on the shelf.

T5 — Wave

Performing a greeting gesture by waving the hand.

Results

The paper reports results for (i) action and task recognition under LOSO cross-validation, (ii) commit decision quality under sensor-dropout scenarios (S0–S6), and (iii) system-level performance with Kinova Gen3 HIL supervisor ablations. Metrics include action/task balanced accuracy and AUPRC, plus decision-quality metrics (FP/1k REST, flaps/min, and success-to-commit coverage), with paired statistical tests and ablations.

Dataset

We release a synchronized tri-modal dataset consisting of EEG (Enobio-8, 8 channels @ 250 Hz), EMG (PLUX, 4 channels), and eye tracking (Pupil Labs Neon) recorded via iMotions/LSL during five ICF-aligned ADL tasks plus REST. Evaluation follows subject-independent leave-one-subject-out (LOSO) splits and sensor-dropout scenarios (S0–S6).

  • T1 (Clock): reaching toward a tabletop clock and grasping it.
  • T2 (Bottle): picking up a dropped apple juice bottle and placing it back onto the table.
  • T3 (Fan): pressing the fan power button to toggle it on and off.
  • T4 (Plant): adjusting or repositioning a plant on the shelf.
  • T5 (Wave): performing a greeting gesture by waving the hand.

BibTeX

If you use this work, please cite:

@inproceedings{sultan2026neurocommitssm,
  title     = {NeuroCommitSSM: Decision-Centric Shared Autonomy for Safe Assistive Manipulation via EEG--EMG--ET Commit Readiness},
  author    = {Sultan, Tipu and Sangani, Param and Cool, Kody and Sikorski, Pascal and Liu, Guangping and Akbarpour, Hadi and Babaiasl, Madi},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  year      = {2026},
  note      = {Under review}
}