EEG-Driven Intention Decoding: Offline Deep Learning Benchmarking on a Robotic Rover

This study establishes a reproducible offline deep learning benchmark for EEG-driven robotic rover control, demonstrating that the ShallowConvNet model outperforms other architectures in decoding user driving intentions across real-time and predictive horizons.

The Gist

Imagine you are driving a remote-controlled car, but instead of using your hands on a joystick, you are using only your brain. That is the dream behind Brain-Computer Interfaces (BCIs). But here's the catch: reading a brain is like trying to hear a single whisper in a crowded, windy stadium. The signals are messy, and the car needs to know exactly what you want before you even finish thinking it.

This paper is a report card on a new experiment that tried to make this dream a reality using a real robot rover outdoors, rather than just a video game.

The Setup: A Brain-Driven Road Trip

The researchers gathered 12 volunteers and sent them on a "road trip" with a 4-wheel-drive robot rover. The volunteers sat in a room, looking at a screen showing what the robot saw (like a first-person video game).

Instead of touching a controller, the volunteers had to think about moving the robot. They were given five specific commands to "think" about:

• Go Forward
• Go Backward
• Turn Left
• Turn Right
• Stop

To capture their thoughts, the volunteers wore a special cap with 16 sensors (like tiny microphones) that listened to their brainwaves (EEG). The robot moved along a real outdoor path, and the researchers recorded everything.

The Big Challenge: Predicting the Future

The tricky part of this experiment was timing.

• The "Now" (0ms): Can the robot know you want to turn right now?
• The "Future" (300ms+): Can the robot guess what you are about to do before you actually do it?

Think of it like a dance partner. A good partner doesn't just follow your moves; they anticipate them. If you lean slightly to the left, they start stepping left before you fully commit. The researchers wanted to see if a computer could be that good dance partner, predicting the driver's intent up to one second in advance.

The Race: 11 AI Coaches

To figure out the best way to decode these brain signals, the researchers didn't just use one method. They lined up 11 different Artificial Intelligence (AI) models to compete. You can think of these models as 11 different coaches trying to teach a computer how to understand the brain:

1. The CNNs (Convolutional Neural Networks): Think of these as detectives who look for specific patterns and shapes in the brainwaves. They are like a magnifying glass looking for clues.
2. The RNNs (Recurrent Neural Networks): These are storytellers. They look at the brainwaves as a sequence of events, remembering what happened a split second ago to understand what's happening now.
3. The Transformers: These are super-spreadsheet analysts. They look at the whole picture at once, trying to find connections between every single part of the brain signal simultaneously.

The Results: Who Won the Race?

After running the data through all 11 coaches, here is what happened:

• The Champion: The winner was a model called ShallowConvNet.

  • The Analogy: Imagine a Swiss Army Knife. It's not the biggest, most complex tool, but it's lightweight, efficient, and gets the job done perfectly. It didn't try to overthink the problem; it just found the right patterns quickly.
  • The Score: It correctly guessed the driver's intent about 67% of the time when looking at the "now," and still managed 66% accuracy when trying to predict the future (300ms ahead).

• The Runners-Up: Other "detective" models (like EEGNet) did well too. The "storyteller" models (like GRU) were also strong, especially for short-term predictions.

• The Losers: Surprisingly, the most complex models (like the "super-spreadsheet" Transformers and very deep networks) didn't do as well.

  • The Analogy: It's like bringing a tank to a bicycle race. These models were too heavy and needed too much data to learn. Since the brain data from just 12 people wasn't "big data" enough, these complex models got confused and overcomplicated things.

Why This Matters

This study is a huge step forward for three reasons:

1. Real World vs. Video Game: Most previous studies were done in labs with fake robots. This one used a real robot on a real path with real outdoor noise (wind, light, movement). It proved that brain-control can work outside the lab.
2. The "Future Sight": The fact that the AI could predict the driver's move 300 milliseconds (0.3 seconds) in advance is huge. In the real world, that split second is the difference between a smooth turn and a crash. It means the robot can start turning before the driver fully decides to, making the control feel natural and fast.
3. Simplicity Wins: The study showed that you don't need a super-complex, heavy AI to read a brain. A simple, well-designed model works better and is faster to run on a robot.

The Bottom Line

This paper is like a blueprint for the future of "mind-controlled" vehicles. It tells us that while the technology is still being tuned, we are getting closer to a world where you can just think "turn left," and a robot will know exactly what to do—even before your hand moves. The secret sauce? Keep the AI simple, keep it fast, and train it on real-world data.

Technical Summary

1. Problem Statement

Brain-Computer Interfaces (BCIs) offer a hands-free control modality for mobile robotics, yet decoding user intent in real-world navigation remains a significant challenge. Existing literature suffers from three primary limitations:

• Simulated Environments: Most studies rely on virtual simulators or controlled indoor labs, lacking the complexity of real-world outdoor navigation.
• Limited Command Sets: Prior work often focuses on binary decisions (e.g., stop/go) or single maneuvers, rather than continuous multi-command decoding (forward, reverse, left, right, stop).
• Lack of Predictive Decoding: Few studies explore "anticipatory" decoding (predicting intent before the physical action occurs) across future time horizons, nor do they systematically compare diverse Deep Learning (DL) architectures under consistent protocols.

This study addresses these gaps by establishing a reproducible benchmark for offline intention decoding during real-world robotic rover operation, comparing 11 state-of-the-art DL models across multiple prediction horizons.

2. Methodology

Experimental Setup

• Participants: 12 healthy adults (6 male, 6 female) performed 10 sessions each.
• Task: Participants remotely operated a 4WD Rover Pro along a predefined outdoor route using an Xbox controller.
• Commands: Five navigational commands were executed: Forward, Reverse, Left, Right, and Stop.
• Hardware:
  • Robot: Equipped with a ZED stereo camera, GNSS, IMU, and Lux sensors (though only EEG was used for training).
  • EEG Acquisition: A 16-channel OpenBCI Cyton + Daisy system with a gel-free cap (10–20 system), sampling at 125 Hz.
  • Synchronization: All data streams (EEG, controller, robot sensors) were synchronized via ROS (Robot Operating System) with sub-millisecond precision.

Data Labeling and Preprocessing

• Labeling Strategy: Joystick inputs (linear/angular velocity) were thresholded to assign discrete labels. Crucially, labels were shifted to create temporal prediction horizons:
  • Action ($\Delta = 0$ ms): Real-time decoding.
  • Intention ($\Delta \in [300, 1000]$ ms): Anticipatory decoding (predicting the command 300ms to 1s in advance).
• Preprocessing Pipeline:
  • High-pass filtering (1 Hz) to remove drift.
  • Adaptive notch filtering (50 Hz) for power line noise.
  • Robust average referencing and interpolation of noisy channels.
  • Z-score normalization.
• Splitting Strategy: A temporal-stratified, label-stratified approach was used to prevent data leakage. Data was grouped by label, sorted chronologically, split into 100 temporal chunks, and then divided 70/30 (Train/Test) within each chunk. This preserved the temporal integrity of the EEG signals.

Deep Learning Benchmarking

Eleven models from three architectural families were benchmarked:

1. Convolutional Neural Networks (CNN): EEGNet, DeepConvNet, ShallowConvNet, STNet, TSCeption, CCNN, CNN1D.
2. Recurrent Neural Networks (RNN): LSTM, GRU.
3. Transformers: ViT (Vision Transformer), EEGConformer.

All models were trained using a within-subject, within-session approach with weighted cross-entropy loss to handle class imbalance.

3. Key Contributions

1. Real-World BCV Benchmark: The first study to combine outdoor robotic rover navigation with synchronized EEG and multi-command decoding (5 classes) under realistic conditions.
2. Predictive Decoding Framework: Introduction of a multi-horizon labeling strategy ($\Delta = 0$ to $1000$ ms) to evaluate both immediate action and anticipatory intention.
3. Systematic DL Comparison: A rigorous comparison of 11 diverse architectures (CNN, RNN, Transformer) using a consistent preprocessing and evaluation pipeline, establishing new performance baselines.
4. Robust Evaluation Protocol: Implementation of a temporal-stratified splitting method to mitigate temporal leakage and ensure metrics reflect real-world generalization.

4. Key Results

• Top Performer: ShallowConvNet achieved the highest performance across all horizons.
  • At $\Delta = 0$ ms (Action): F1-score 67% (Accuracy 83%).
  • At $\Delta = 300$ ms (Intention): F1-score 66% (Accuracy 82%).
  • It maintained robust performance (F1 > 60%) up to $\Delta = 900$ ms.
• Runner-ups:
  • EEGNet and CNN1D showed competitive results (F1 ~62-64%).
  • GRU (RNN) performed well (F1 ~62-63%) but showed higher variability across subjects compared to CNNs.
  • EEGConformer (Transformer hybrid) was stable (F1 ~60-61%) but did not outperform compact CNNs.
• Underperformers:
  • DeepConvNet suffered from overfitting (F1 ~24%), likely due to over-parameterization on the limited dataset.
  • ViT (pure Transformer) struggled (F1 ~52%), highlighting the data-hunger of attention mechanisms in small-sample EEG contexts.
• Temporal Robustness: A gradual decline in F1-score was observed as the prediction horizon increased, but top models retained >60% F1-score up to 900ms, proving the feasibility of anticipatory control.

5. Significance and Conclusion

This study demonstrates that compact CNNs (specifically ShallowConvNet) offer the best trade-off between performance, robustness, and efficiency for real-time brain-robot control, outperforming complex RNNs and Transformers in this specific context.

• Practical Implication: The results confirm that EEG-based systems can reliably anticipate user driving intentions up to 0.9 seconds in advance. This latency is critical for shared autonomy, allowing robotic systems to react proactively to human intent rather than just reacting to executed movements.
• Future Directions: The authors propose extending this framework to online real-time experiments, cross-subject generalization, and multimodal fusion. The benchmark serves as a foundational reference for developing neuroadaptive control systems for intelligent vehicles and assistive robotics.