SPOT: Single-Shot Positioning via Trainable Near-Field Rainbow Beamforming

This paper proposes SPOT, an end-to-end deep learning framework that jointly optimizes phase-time array beamforming and user position estimation to achieve high-accuracy single-shot localization with significantly reduced overhead compared to existing methods.

The Gist

Imagine you are in a large, dark warehouse trying to find a group of people hiding in different corners. You have a special flashlight, but instead of just shining a single beam, this flashlight can split its light into a rainbow.

In this paper, the researchers (Yeyue Cai, Jianhua Mo, and Meixia Tao) propose a new, super-smart way to use this "rainbow flashlight" to find people quickly and accurately, without needing them to shout back a long list of details. They call their method SPOT.

Here is the breakdown of how it works, using simple analogies:

1. The Problem: The Old Way is Slow and Clunky

Traditionally, to find someone in a room, a radar or base station has to play a game of "hot and cold."

• The Old Method: The base station shines a beam in one direction, waits for a reply, then moves the beam slightly, waits again, and repeats this dozens of times. It's like sweeping a flashlight across a room slowly, checking every inch.
• The Feedback Problem: Sometimes, the people being found have to send back a huge amount of data (like a full photo of what they see) so the base station can figure out where they are. This clogs up the network and takes a long time.

2. The Solution: The "Rainbow Beam"

The researchers use a special hardware setup called a Phase-Time Array (PTA). Think of this as a flashlight that doesn't just shine one color at one angle.

• The Rainbow Effect: When you turn it on, it sends out a "rainbow" of light.
  • Red light (low frequency) goes to the left.
  • Blue light (high frequency) goes to the right.
  • Green light goes straight ahead.
• The Magic: Because different colors go in different directions, the system can scan the entire room instantly with just one flash. If a person is standing on the left, they only see the red light. If they are on the right, they only see the blue light.

3. The Innovation: Teaching the Flashlight to "Think"

This is where the paper gets really clever.

• Old Way: Engineers used to manually calculate how to set the flashlight's lenses to make the rainbow. They guessed the settings based on math formulas.
• The SPOT Way: The researchers treated the flashlight's settings like a video game character's stats. They used Artificial Intelligence (Deep Learning) to "train" the flashlight.
  • The AI tried millions of different settings.
  • It learned that to find people best, the rainbow shouldn't just be a standard rainbow; it should be a customized rainbow that highlights exactly where people are likely to be.
  • The AI "learned" how to twist the light waves so that the signal is strongest exactly where the user is standing.

4. The "One-Shot" Trick

This is the biggest efficiency gain.

• The Old Game: "I shine a beam. You tell me what you see. I shine another beam. You tell me what you see." (This takes two or more rounds).
• The SPOT Game:
  1. One Flash: The base station sends the single, smart, AI-optimized rainbow beam.
  2. One Tiny Reply: The user doesn't need to send a photo or a long message. They just send back two tiny numbers:
     • "I saw the Red light." (Which subcarrier/frequency was strongest).
     • "It was very bright." (How strong the signal was).
  3. The Guess: The base station looks at those two numbers and instantly knows: "Ah, Red means they are at a 30-degree angle, and the brightness means they are 50 meters away."

5. Why is this better?

• Speed: It finds everyone in a single shot instead of sweeping the room slowly.
• Data Savings: Instead of sending back a whole file, the user sends a message as small as a text emoji. This saves massive amounts of network traffic.
• Accuracy: Because the AI "learned" the perfect beam pattern, it doesn't get confused by the distance or the angle. It works great whether the person is close (near-field) or far away.

The Bottom Line

Imagine trying to find a friend in a crowded stadium.

• The Old Way: You shout "Where are you?" and wait for them to describe their seat number, row, and section. Then you shout again to confirm.
• The SPOT Way: You flash a giant, colorful laser grid over the whole stadium. Your friend just raises their hand and says, "I'm under the Blue light, and it's super bright." You instantly know exactly where they are.

The paper proves that by using AI to design the "laser grid" (the beam) and asking for a very simple answer, we can locate people faster, cheaper, and more accurately than ever before. This is a huge step forward for 6G networks and future smart cities.

Technical Summary

1. Problem Statement

The paper addresses the challenge of user positioning in wideband Integrated Sensing and Communication (ISAC) systems using Phase-Time Arrays (PTAs).

• Context: PTAs combine Phase Shifters (PSs) and True-Time Delays (TTDs) to generate "rainbow beams," where different OFDM subcarriers point to different spatial directions. This allows for rapid scanning compared to conventional phased arrays.
• Limitations of Existing Methods:
  • High Overhead: Most existing schemes require multiple downlink transmission rounds (e.g., one for angle, one for distance) or extensive uplink feedback (e.g., full received signal vectors), leading to significant latency and signaling overhead.
  • Suboptimal Beam Design: Current approaches often rely on heuristic or analytical designs for PS/TTD coefficients. These fixed designs often fail to optimize the mapping between frequency, angle, and distance, especially in the near-field where spherical wavefronts dominate.
  • Error Propagation: Multi-stage methods (estimating angle first, then distance) suffer from error propagation, where initial angular errors degrade distance estimation.
  • Resource Waste: Standard rainbow beams often cover wide sectors (e.g., $-90^\circ$ to $90^\circ$) regardless of the actual user distribution, causing unnecessary resource consumption.

2. Methodology: The SPOT Scheme

The authors propose SPOT (Single-Shot Positioning via Trainable near-field rainbow beamforming), an end-to-end deep learning framework that jointly optimizes beamforming and position estimation.

A. System Model

• Architecture: A Base Station (BS) with a Uniform Linear Array (ULA) connected to a single RF chain via $N$ PSs and $N$ TTDs.
• Channel: Near-field spherical wave model is assumed for $K$ users.
• Process: The BS transmits a single downlink rainbow beam. Each user identifies the subcarrier with maximum received power and feeds back only two values: the quantized peak power and the subcarrier index.

B. Network Architecture (Two Jointly Trained Modules)

The system is modeled as a differentiable pipeline optimized to minimize 2D localization Root Mean Square Error (RMSE).

1. Module 1: Trainable Rainbow Beamformer

   • Innovation: Instead of fixed heuristics, the PS coefficients ($\Phi$) and TTD coefficients ($T$) are treated as trainable network parameters ($\Theta_\Phi, \Theta_T$).
   • Function: The network learns to synthesize task-oriented beams that maximize the discriminability of user positions.
   • Differentiability: To enable backpropagation through the "arg max" operation (selecting the peak subcarrier), the authors use a temperature-scaled softmax during training to approximate the hard decision with a soft, differentiable weight.

2. Module 2: User Localization Estimator

   • Input: The soft subcarrier index ($m^\star_k$) and the soft peak power ($p^\star_k$).
   • Structure: A lightweight Fully Connected (FC) Deep Neural Network (DNN) with 4 layers.
   • Output: Estimates the user's angle ($\hat{\phi}_k$) and distance ($\hat{r}_k$).
   • Constraints: Activation functions (tanh and Softplus) are used to ensure physical constraints (e.g., angle within $[-60^\circ, 60^\circ]$, distance $>0$).

C. Deployment Strategy

• Inference: During actual deployment, the soft selection is replaced by a hard decision (true arg max).
• Quantization: The peak power is uniformly quantized (e.g., 8 bits) to reduce feedback size.
• Parameter Projection: Learned PS/TTD values are projected onto physical ranges (modulo operations) to ensure hardware feasibility without affecting performance due to periodicity.

3. Key Contributions

1. Single-Shot Mechanism: The scheme achieves 2D positioning (angle and range) using only one downlink transmission and a single uplink feedback message containing just two scalar values (index + quantized power).
2. End-to-End Joint Optimization: By treating PS and TTD coefficients as learnable parameters, the system generates "task-oriented" rainbow beams specifically optimized for the localization task, outperforming heuristic designs.
3. Lightweight Feedback & Computation:
   • Reduces feedback overhead by an order of magnitude compared to schemes requiring full signal vectors.
   • Achieves extremely low computational complexity ($0.035$ M FLOPs) compared to ConvNeXt or RaiNet baselines.
4. Robustness: The learned beams dynamically adapt their power-distance profiles to the user range, maintaining accuracy from near-field (5m) to far-field (300m).

4. Numerical Results

Simulations were conducted with a 28 GHz carrier, 380 MHz bandwidth, and 256 antennas. Users were distributed between 5m and 300m.

• Accuracy (RMSE):
  • SPOT consistently outperforms baselines (CBS, ConvNeXt, Distance-dependent CBS, RaiNet).
  • Compared to a fixed-beam baseline with a trained estimator, SPOT reduces 2D RMSE by ~40.9%.
  • Compared to the original RaiNet (which uses full power vectors), SPOT achieves comparable or better accuracy with significantly lower complexity.
• Overhead Reduction:
  • Feedback: SPOT requires only 19 bits per user (11-bit index + 8-bit power), whereas RaiNet requires ~50,000 bits (full vector) and ConvNeXt requires ~200,000 bits.
  • Latency: SPOT requires only 1 downlink round, whereas CBS and ConvNeXt require 2 rounds.
• Quantization Robustness: The system maintains high accuracy even with low-bit quantization (e.g., 8-bit power feedback yields performance nearly identical to unquantized feedback).
• Beam Patterns: The learned SPOT beams exhibit a monotonic power-distance relationship across angles, acting as an effective spatial encoder, whereas conventional CBS beams suffer from resolution loss at long distances.

5. Significance

This paper represents a significant advancement in ISAC and Near-Field Localization:

• Paradigm Shift: It moves away from analytical beam design toward data-driven, learnable beamforming, proving that the physical layer (hardware coefficients) can be optimized directly for the application layer (localization).
• Practicality: By drastically reducing feedback overhead and latency, SPOT makes high-precision, wideband near-field positioning feasible for real-time 6G applications where signaling resources are scarce.
• Efficiency: It demonstrates that complex positioning tasks can be solved with minimal computational resources and minimal signaling, making it highly suitable for energy-constrained and high-density networks.