← Paper Digest

Anchoring Motion, Refining Trajectories: Generalizable Continuous Neural Decoding

Fanqi Shen, Junru Chen, Surui Cheng, Jiahe Li, ZhiZhang Yuan, Gaorui Zhang, ZhiYuan Cao, Xiaoquan Shen, Yang Yang

November 2026

TL;DR AMRT decodes continuous hand and cursor trajectories from scalp EEG and intracranial ECoG by first predicting velocity, then correcting the drift of its integral, with a channel adapter for differing electrode layouts. On unseen subjects of the U-Limb EEG dataset it reaches a trajectory correlation of 0.430, against 0.160 for the best baseline.

What they do

Continuous control of cursors, robotic effectors and neuroprostheses needs time-varying kinematics, not discrete movement classes, and decoders that map brain signals straight to position tend to be tied to one subject, electrode layout and workspace. AMRT is one decoder for scalp EEG and intracranial ECoG that predicts local motion first and the full trajectory second. It is tested on unseen trials, unseen subjects and transfer between public datasets.

How it works

  • Channel adapter. EEG channels map onto a reference montage plus a small learnable residual. iEEG electrodes, whose number and placement differ per patient, are treated as an unordered set embedded with their 3D coordinates and pooled by attention.
  • Encoder. A BiLSTM and multi-horizon blocks (temporal convolutions of several widths plus global attention, gated) read a 5 s context.
  • Stage-0, motion anchoring. A query-based action-chunk decoder predicts 1 s of velocity; overlapping chunks are blended and integrated into a coarse trajectory.
  • Stage-1, trajectory refinement. With Stage-0 frozen, a second decoder reads the coarse trajectory and the neural features and predicts gated residual corrections that remove integration drift.

Results

Metric is trial-level trajectory correlation (PCC) on WAY-EEG-GAL (32-channel EEG, grasp-and-lift, 12 subjects), U-Limb (128-channel EEG, 30 upper-limb tasks, 20 subjects) and Joystick-Track (ECoG, 4 patients).

  • Unseen trials, subjects pooled: 0.530 vs. 0.314 for the best baseline on U-Limb, 0.707 vs. 0.464 on Joystick-Track with shared channels, 0.837 vs. 0.813 on WAY-EEG-GAL.
  • Leave-one-subject-out: 0.841, 0.430 and 0.719 on WAY-EEG-GAL, U-Limb and Joystick-Track; the strongest baseline, M3T-Attention, gets 0.785 and 0.160 on the two EEG sets.
  • Pretraining on ECoG Joystick-Track and adapting only Stage-1 on 10% of U-Limb gives 0.454, above the 0.192 of training from scratch on all U-Limb data. Neither EEG dataset transfers to Joystick-Track.
  • Velocity-then-refine beats direct position regression across subjects (U-Limb 0.430 vs. 0.268), but within a single subject direct regression is slightly better (WAY-EEG-GAL 0.934 vs. 0.914).

Limitations

Decoding is offline on recorded data; nothing is controlled in closed loop. Transfer depends on the source-target pairing, and a second source dataset does not reliably help. On WAY-EEG-GAL even an EEG-free trajectory template scores well, so those numbers partly reflect stereotyped movements.

Figures

From Figure 1: The AMRT framework. Fixed-montage EEG or variable-layout iEEG passes through a channel adapter and a n...
From Figure 1: The AMRT framework. Fixed-montage EEG or variable-layout iEEG passes through a channel adapter and a neural encoder (BiLSTM and multi-horizon blocks). Stage-0 predicts velocity chunks that integrate into a coarse trajectory; Stage-1 refines it with gated residual corrections.
From Figure 2: (a) Few-shot adaptation to held-out subjects across calibration sizes. (b) Cross-dataset transfer acro...
From Figure 2: (a) Few-shot adaptation to held-out subjects across calibration sizes. (b) Cross-dataset transfer across target-data fractions. (c) Stage-1-only gains over matched scratch training. (d) Stage-1 input ablation (full, coarse-only, neural-only). (e) Example trajectory reconstructions from plain velocity integration and from AMRT.
Abstract

Brain activity encodes motor information across multiple temporal scales, enabling continuous neural control beyond discrete action recognition. However, existing methods are typically tailored to specific subjects and tasks, with direct neural-to-kinematic mappings that can entangle subject-specific neural variability with task-dependent motion patterns, limiting generalization. To address these limitations, we introduce AMRT, a shared neural decoding framework for heterogeneous electrode configurations and movement paradigms. Motivated by cortical encoding of movement direction and speed, AMRT adopts a two-stage design that anchors local motion through velocity estimation and refines the resulting trajectories over longer horizons, encouraging reuse of local motion tendencies across subjects and tasks. Across three datasets, AMRT generalizes to unseen trials and subjects and demonstrates transfer potential across neural recordings and movement paradigms, highlighting the potential of shared neural representations and anchor-and-refine decoding for generalizable continuous neural control.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.