⚡ electrical engineering

A Tutorial on AI-Empowered Integrated Sensing and Communications

This tutorial paper explores how artificial intelligence, particularly deep learning, can optimize the design of unified waveforms, constellations, and beamformers for Integrated Sensing and Communications (ISAC) systems to effectively balance performance, complexity, and implementation constraints where traditional model-based approaches fall short.

Original authors: Mojtaba Vaezi, Gayan Aruma Baduge, Esa Ollila, Sergiy A. Vorobyov

Published 2026-02-16

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Mojtaba Vaezi, Gayan Aruma Baduge, Esa Ollila, Sergiy A. Vorobyov

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to host a party where you need to do two very different things at the same time: talk to your guests (Communication) and keep an eye on the room to see who is moving or where they are (Sensing/Radar).

Traditionally, you would use two separate tools: a microphone to talk and a security camera to watch. This works, but it's expensive, takes up a lot of space, and uses up a lot of electricity.

Integrated Sensing and Communications (ISAC) is the idea of using one single tool to do both jobs. Think of it like a "smart flashlight" that not only shines light so people can see each other (communication) but also bounces the light off walls to map out the room and detect movement (sensing).

However, there's a catch. The best way to shine a light for talking (which needs to be random and varied to carry complex messages) is often the worst way to shine it for sensing (which needs to be very structured and predictable to bounce off things clearly). It's like trying to write a poem while simultaneously trying to draw a perfect circle; the two goals often fight each other.

This paper is a tutorial (a "how-to" guide) on how to use Artificial Intelligence (AI) to solve this problem. Instead of trying to write complex math equations to figure out the perfect balance, the authors show how AI can "learn" the perfect solution by trial and error, just like a human learns to juggle.

Here is a breakdown of the paper's main ideas using simple analogies:

1. The Problem: The "Juggling Act"

In the old days, engineers tried to design these systems using strict math rules (Model-Based).

The Analogy: Imagine trying to design a new car engine by only using a ruler and a calculator. You can do it, but it takes forever, and if the road conditions change (like a sudden storm or a pothole), your perfect calculation might fail.
The Issue: The math gets too complicated when you try to balance talking and sensing. Sometimes the math is so hard that the computer can't solve it fast enough for real-time use (like driving a car).

2. The Solution: The "Smart Apprentice" (AI)

The paper suggests using Deep Learning (a type of AI) as a "smart apprentice." Instead of giving the AI a rulebook, you let it watch thousands of examples and learn the patterns itself.

The Analogy: Instead of telling the apprentice how to juggle with a manual, you let them practice juggling balls and apples thousands of times. Eventually, they figure out the perfect rhythm to keep everything in the air without you having to explain the physics of gravity.

3. Three "Training Camps" (Case Studies)

The authors tested this AI idea in three specific ways, which they call "Case Studies":

Case Study 1: The "Shape-Shifter" (Waveform Design)
- The Goal: Create a signal shape that is good for both talking and sensing.
- The AI Trick: They used Unsupervised Learning. Imagine the AI is given a lump of clay and told, "Make this clay look like a sphere (for sensing) but also feel like a cube (for communication)." The AI squishes and stretches the clay until it finds a weird, perfect shape that satisfies both requirements better than a human could design.
- Result: The AI found a shape that was almost as good as the "perfect" math solution but took a fraction of the time to compute.
Case Study 2: The "Speed-Runner" (Beamforming)
- The Goal: Directing the signal like a laser beam to specific people while ignoring others.
- The AI Trick: They used Algorithm Unrolling. Imagine a traditional math method is like climbing a mountain step-by-step, checking your map at every single step. It's safe but slow. The AI method is like taking a "shortcut" it learned from experience. It knows exactly which steps to skip to get to the top faster.
- Result: The AI found the best direction for the beam much faster than the traditional method, especially when the environment was messy.
Case Study 3: The "New Alphabet" (Constellation Design)
- The Goal: Creating a new "alphabet" of signals. Traditional alphabets are either great for reading (Communication) or great for bouncing off walls (Sensing), but not both.
- The AI Trick: They used an Autoencoder (a type of AI that learns to compress and then expand data). The AI invented a brand new "alphabet" of signal points that didn't look like the standard squares (QAM) or circles (PSK) we use today. It created a "hybrid" alphabet that was a perfect compromise.
- Result: The AI's new alphabet was better at both talking and sensing than the old standard alphabets.

4. Why This Matters for the Future (6G)

We are moving toward 6G networks (the next generation of mobile internet). The goal of 6G is "Everything is sensed, everything is connected, and everything is intelligent."

The Vision: Your phone won't just call your mom; it will also act as a radar to help your car avoid accidents, or help a drone navigate a city.
The AI Role: Because 6G will be so complex and change so fast, we can't rely on old math rules. We need AI to adapt in real-time, just like a human driver adapts to traffic.

5. The Catch (Limitations)

The paper is honest about the downsides:

The "Training" Cost: The AI needs a lot of data to learn. If the environment changes drastically (like a sudden earthquake or a new type of interference), the AI might get confused and need to be "re-trained."
The "Black Box": Sometimes, even the AI doesn't know why it made a decision. It just knows it works. This can be scary for safety-critical systems.

Summary

This paper is a guide on how to stop trying to force a square peg into a round hole with complex math. Instead, it shows how to use AI as a flexible, learning tool that can invent new ways to combine talking and sensing. It's like teaching a robot to be both a radio host and a security guard simultaneously, using a single, super-efficient tool that learns from experience rather than a rigid rulebook.

1. Problem Statement

Integrated Sensing and Communications (ISAC) aims to unify radar sensing and wireless communication into a single system to address spectrum scarcity, hardware costs, and energy inefficiency. However, ISAC faces fundamental trade-offs:

Conflicting Objectives: Communication prioritizes high data rates (maximizing Mutual Information, MI), often requiring random, high-entropy signals (e.g., Gaussian inputs). Sensing prioritizes accurate parameter estimation (minimizing Mean Squared Error, MSE), often favoring deterministic, constant-envelope signals (e.g., BPSK/PSK).
Optimization Complexity: Designing unified waveforms, beamformers, and constellations requires solving complex, non-convex, multi-objective optimization problems.
Model Limitations: Traditional model-based analytical approaches often fail due to:
- Incomplete or inadequate system models (e.g., imperfect Channel State Information, hardware impairments).
- Prohibitive computational complexity for real-time implementation (e.g., Semi-Definite Relaxation (SDR) scales poorly with antenna counts).
- Inability to adapt to dynamic, unpredictable environments.

The paper posits that Artificial Intelligence (AI), particularly Deep Learning (DL), offers a data-driven alternative to overcome these limitations by learning implicit channel features and optimizing complex trade-offs without explicit analytical models.

2. Methodology

The paper structures its approach around Multi-Objective Optimization (MOO) and various AI paradigms:

A. Optimization Frameworks

Traditional Methods: Discusses Weighted Sum, $\epsilon$ -Constraint, and Pareto Optimization methods. These often struggle with non-convex Pareto fronts and high computational costs.
Learning-Based Approaches:
- Single-Task Learning: Optimizes a joint loss function (weighted sum of sensing and communication metrics) using a single neural network.
- Multi-Task Learning (MTL): Uses shared layers for common features and task-specific layers for distinct outputs (e.g., channel estimation and target detection).
- Multi-Objective Learning (MOL): Balances conflicting objectives within a single task (e.g., maximizing rate while minimizing latency).
- Model-Based Learning:
  - Plug-and-Play (PnP): Replaces complex algorithmic steps (like denoisers) with neural networks.
  - Algorithm Unrolling: Unfolds iterative optimization algorithms (e.g., Projected Gradient Ascent) into neural network layers, learning hyperparameters (like step sizes) to escape local minima and converge faster.
- End-to-End (E2E) Learning: Treats the entire transmitter-receiver chain as an autoencoder, jointly optimizing modulation, coding, and beamforming.

B. Key Design Areas

The tutorial focuses on three core physical-layer designs:

Waveform Design: Creating unified signals that satisfy both radar correlation properties and communication spectral efficiency.
Beamforming Design: Jointly optimizing analog and digital precoders to direct energy for sensing and communication simultaneously.
Constellation Design: Shaping signal constellations to balance the randomness needed for data transmission and the structure needed for sensing accuracy.

3. Key Contributions

The paper provides a comprehensive tutorial and three specific case studies demonstrating AI's efficacy:

Case Study I: Unsupervised Learning for Waveform Design
- Method: A Deep Neural Network (DNN) with a custom Lambda layer to enforce power constraints. It uses an unsupervised loss function combining Multi-User Interference (MUI) minimization (communication) and waveform mismatch minimization (sensing).
- Result: The AI-designed waveform achieves a sum rate within 0.88%–1.31% of the optimal analytical solution (SDR-based) but with significantly lower computational complexity. It effectively generates unified beampatterns for both omnidirectional and directional sensing.
Case Study II: Algorithm Unrolling for Hybrid Beamforming
- Method: Unrolls a Projected Gradient Ascent (PGA) algorithm into a neural network to learn optimal step sizes for hybrid analog/digital beamforming.
- Result: The "Unrolled PGA" converges much faster than standard PGA and achieves a higher achievable sum rate across all SNR levels. It effectively escapes local optima that trap traditional iterative algorithms.
Case Study III: Autoencoder-Based Constellation Design
- Method: An E2E autoencoder architecture where a shared encoder generates a signal for both a communication decoder and a radar detector. The loss function balances Bit Error Rate (BER) and Probability of Detection ( $P_D$ ).
- Result: The system learns novel constellations that adapt to the trade-off parameter ( $\eta$ $η$ ).
  - $\eta \approx 0$ (Comm-centric): Learns QAM-like structures.
  - $\eta \approx 1$ (Sensing-centric): Learns PSK-like structures.
  - Balanced $\eta$ : Learns multi-ring constellations that outperform traditional 32-QAM and 32-PSK in joint performance metrics.

4. Results and Performance Evaluation

Complexity vs. Performance: AI-based designs consistently offer a "sweet spot," providing near-optimal performance (often <2% gap from analytical bounds) with drastically reduced computational overhead, making them suitable for real-time 6G applications.
Robustness: AI models demonstrate better adaptability to hardware impairments and non-linearities compared to convex optimization methods.
Generalization Limitations: The paper highlights a critical limitation: AI models trained on specific channel distributions or topologies (e.g., specific number of users) suffer performance degradation when faced with channel aging (user mobility) or topology changes (adding/removing users). Retraining or domain adaptation is required.

5. Significance and Future Directions

Significance:

Paradigm Shift: Moves ISAC design from rigid, model-based analytical solutions to flexible, data-driven learning frameworks.
6G Enabler: Provides the necessary low-complexity, high-adaptability tools required for the "AI-native" air interface of 6G.
Unified Design: Demonstrates that a single signal can effectively serve dual purposes, resolving the historical separation of radar and comms.

Future Research Directions Identified:

End-to-End ISAC: Extending E2E learning to multi-cell scenarios with complex interference (self-interference, inter-user interference).
Transformer Networks: Leveraging Transformers for their ability to model long-range dependencies and fuse multi-modal data (vision, LiDAR, radar) for dynamic beamforming.
Digital Twins: Using digital replicas of physical environments to train and validate AI models before real-world deployment.
FR3 Spectrum: Adapting learning algorithms for the 7–24 GHz "mid-band" spectrum, which has unique propagation characteristics (moderate sparsity, near-field effects) not covered by current sub-6 GHz or mmWave models.

In conclusion, this tutorial establishes that while traditional optimization provides theoretical bounds, AI-empowered designs are the practical path forward for realizing efficient, robust, and unified ISAC systems in next-generation networks.