Beyond Amplitude: Channel State Information Phase-Aware Deep Fusion for Robotic Activity Recognition

This paper introduces GF-BiLSTM, a two-stream gated fusion network that systematically demonstrates the critical role of incorporating Wi-Fi Channel State Information phase data alongside amplitude to significantly improve the accuracy and cross-speed robustness of robotic activity recognition.

The Gist

Imagine you are trying to guess what a robot is doing just by listening to the Wi-Fi signals bouncing around the room. It's like trying to figure out if someone is dancing, waving, or holding still just by watching the ripples in a pond when they drop a stone.

This paper is about a new, smarter way to "listen" to those Wi-Fi ripples to recognize what a robotic arm is doing, even if the robot moves at different speeds.

Here is the breakdown of their discovery, using simple analogies:

1. The Problem: Only Listening to the Volume

For a long time, scientists used Wi-Fi to track movement by looking at the Amplitude (the strength or "volume" of the signal).

• The Analogy: Imagine trying to guess what a singer is doing just by looking at a volume meter. If the singer gets louder, the meter goes up. If they whisper, it goes down.
• The Flaw: This is okay, but it misses the nuance. It doesn't tell you how the sound is changing, just how loud it is. Also, in a room full of echoes (like a robot moving), the volume meter gets confused easily.

2. The Secret Ingredient: The "Phase"

The researchers realized they were ignoring the Phase.

• The Analogy: If Amplitude is the volume, Phase is the timing or the rhythm. It tells you exactly when the sound wave hits your ear relative to when it left the source.
• The Catch: Raw Phase data is messy. It's like trying to read a clock that keeps jumping forward and backward randomly because the battery is weak. The hardware makes the signal "jitter."

3. The Solution: The "Gatekeeper" Network (GF-BiLSTM)

The team built a new AI model called GateFusion-BiLSTM. Think of this model as a very smart Traffic Controller or a Conductor.

• Two Lanes of Traffic: The model has two lanes. One lane looks at the Volume (Amplitude), and the other looks at the Rhythm (Phase).
• Cleaning the Rhythm: Before the Rhythm lane can be used, the model performs "Sanitization." It's like a spell-checker that fixes the jittery clock so the numbers make sense.
• The Magic Gate: This is the best part. The model has a "gate" that decides, millisecond by millisecond, which lane to trust.
  • If the robot is moving smoothly, the gate might say, "The Rhythm lane is clear, let's use that!"
  • If the signal gets noisy, the gate says, "The Rhythm is too messy right now, ignore it and rely on the Volume lane."
  • It learns to blend the two perfectly, like a DJ mixing two tracks so the beat never skips.

4. The Experiment: The "Speed Test"

To test this, they used a dataset where a robot arm performed 8 different actions (like drawing a triangle or an arc) at three different speeds: Slow, Medium, and Fast.

They used a strict test called "Leave-One-Velocity-Out."

• The Analogy: Imagine you teach a student to drive using only a slow car and a fast car. Then, you put them in a medium-speed car they've never seen before and ask them to drive. Can they do it?
• The Result: Most old models failed this test. They were too rigid. But the new GateFusion model passed with flying colors. Because it learned to listen to both volume and rhythm, it could recognize the robot's actions even when the speed changed completely.

5. The Trade-off: Is it worth the extra work?

The researchers found that "cleaning" the Phase data (Sanitization) made the model slightly more accurate, but it took a lot more computer power to do the cleaning.

• The Verdict: They found a "sweet spot." Using the raw "unwrapped" Phase (just fixing the jumps, not the deep cleaning) gave them 95% of the benefit with much less computing cost. It was the most efficient way to get the job done.

The Big Takeaway

This paper proves that to truly understand what a robot is doing using Wi-Fi, you can't just listen to the volume of the signal. You have to listen to the rhythm (Phase) too.

By building a system that knows when to trust the rhythm and when to trust the volume, they created a robot-sensing system that is much more robust, accurate, and capable of handling real-world chaos. It's the difference between a security guard who only hears footsteps and one who also understands the gait of the person walking.

Technical Summary

Here is a detailed technical summary of the paper "Beyond Amplitude: Channel State Information Phase-Aware Deep Fusion for Robotic Activity Recognition."

1. Problem Statement

While Wi-Fi Channel State Information (CSI) has emerged as a promising non-line-of-sight (NLOS) modality for activity recognition, existing research predominantly relies on CSI amplitude while largely ignoring phase information. This is primarily because raw phase data is corrupted by hardware-induced timing and frequency offsets, making it noisy and difficult to utilize.

The specific gap addressed in this paper is the lack of systematic exploration into using CSI phase for robotic arm activity recognition. Unlike human activities, robotic movements are precise and repetitive, potentially offering unique phase signatures. The authors investigate whether incorporating phase (after appropriate preprocessing) can improve recognition accuracy and, crucially, the model's robustness to variations in execution speed (velocity).

2. Methodology

A. Data Preprocessing

The raw complex CSI ($H$) is decomposed into Amplitude ($A$) and Phase ($\Phi$). The phase undergoes a rigorous two-step preprocessing pipeline to remove hardware artifacts:

1. Temporal Unwrapping: Resolves the $(-\pi, \pi]$ wrapping discontinuity to create continuous phase trajectories.
2. Linear Sanitization: Removes packet-wise linear trends (caused by residual synchronization errors) using a least-squares fit across subcarriers.
   • Note: The paper evaluates both "Unwrapped" phase and "Sanitized" phase to balance accuracy against computational cost.

B. Proposed Architecture: GateFusion-BiLSTM (GF-BiLSTM)

The core contribution is a two-stream gated fusion network designed to adaptively integrate amplitude and phase features.

• Dual Streams: The amplitude and phase inputs are processed separately through parallel Bidirectional Long Short-Term Memory (BiLSTM) networks.
• Normalization: Each stream undergoes layer normalization across subcarriers.
• Gated Fusion Mechanism: Instead of simple concatenation, the model employs a learned gating mechanism at each time step.
  • A gate vector $g_t$ is computed via a sigmoid function based on the concatenated features of both streams.
  • The final fused feature $z_t$ is a weighted sum: $z_t = g_t \odot u^A_t + (1 - g_t) \odot u^P_t$.
  • This allows the model to dynamically down-weight noisy phase windows and rely more on amplitude when necessary, and vice versa.
• Classification: The fused sequence passes through deeper BiLSTM layers, temporal average pooling, and a Multi-Layer Perceptron (MLP) head for action classification.
• Regularization: A "modality dropout" is applied during training, randomly masking one entire stream (amplitude or phase) to force the model to learn robust representations from both modalities independently.

C. Evaluation Protocol

• Dataset: RoboFiSense benchmark (8 robotic arm actions: Arc, Elbow, Rectangle, etc.).
• Protocol: Leave-One-Velocity-Out (LOVO). Models are trained on two velocity levels (e.g., Low and Medium) and tested on the unseen third (e.g., High). This rigorously tests generalization across motion speeds.
• Baselines: CNN, LSTM, BiLSTM, Vision Transformer (ViT), and Bidirectional ViT Concatenated (BiVTC).

3. Key Contributions

1. Systematic Phase Exploration: This is the first work to systematically evaluate the role of CSI phase (raw, unwrapped, and sanitized) specifically for robotic activity recognition, establishing it as a critical complementary cue rather than a substitute for amplitude.
2. GF-BiLSTM Architecture: Introduction of a novel gated-fusion architecture that learns to adaptively weight amplitude and phase features per time step, significantly outperforming standard fusion methods.
3. Cross-Speed Robustness: Demonstration that phase-aware models generalize significantly better to unseen execution speeds compared to amplitude-only models.
4. Efficiency Analysis: A detailed trade-off analysis showing that while linear sanitization yields the highest accuracy, temporal unwrapping alone offers a superior accuracy-to-computational-cost ratio.

4. Experimental Results

The study was conducted on the RoboFiSense dataset with 8 classes and 3 velocity levels.

• Input Configurations:

  • Phase Only: Performed worst (lowest accuracy) due to high noise and lack of amplitude context.
  • Amplitude Only: Provided a stable baseline but struggled with speed variations.
  • Amplitude + Unwrapped Phase: Significant improvement over single modalities.
  • Amplitude + Sanitized Phase: Achieved the highest absolute accuracy but at a high computational cost.

• Model Performance (LOVO Protocol):

  • The proposed GF-BiLSTM achieved the best performance across all configurations.
  • Top Result: GF-BiLSTM with Amplitude + Sanitized Phase achieved 96.11% accuracy (Training on V1&V2, Testing on V3).
  • Comparison: GF-BiLSTM consistently outperformed strong baselines like BiVTC (93.85%) and BiLSTM (91.11%) in the same configuration.
  • Generalization: The gap between GF-BiLSTM and baselines widened in cross-speed scenarios, proving the gated fusion mechanism effectively handles speed-induced distribution shifts.

• Computational Cost:

  • Amplitude-only preprocessing: ~1.64 ms/sample.
  • Amplitude + Unwrapped Phase: 12.52 ms/sample (7.7x cost).
  • Amplitude + Sanitized Phase: 78.41 ms/sample (47.9x cost).
  • Conclusion: The authors recommend the Unwrapped variant for real-time applications as it offers the best balance, with sanitization adding marginal accuracy gains for a massive computational penalty.

5. Significance

This paper fundamentally shifts the paradigm in Wi-Fi sensing for robotics by proving that phase information is indispensable for high-precision activity recognition, particularly when dealing with variable execution speeds.

• Technical Impact: It validates that deep learning models can effectively learn to "trust" or "distrust" phase data dynamically via gating mechanisms, overcoming the historical barrier of phase noise.
• Practical Application: The findings enable more robust, privacy-preserving robotic monitoring systems that do not require Line-of-Sight (unlike cameras/LiDAR) and can adapt to robots operating at different speeds without retraining.
• Future Direction: The work opens avenues for complex-valued neural networks and polar coordinate inputs in Wi-Fi sensing.