TL;DR
Ultra-CAN is a lightweight, multi-task neural receiver that allows drones to perform both secure telemetry and radar ranging using a single chaotic waveform. By sharing a compact feature extractor, it achieves state-of-the-art performance on low-power microcontrollers (like the ESP32-S3), providing a robust, jam-resistant solution for the Internet of Drones (IoD).
What Problem Does This Paper Address?
Drones operating in swarms face a difficult balancing act. They need secure, jam-resistant communication (telemetry) and the ability to sense their surroundings (radar) to avoid collisions. However, micro-UAVs are bound by strict SWaP-C (Size, Weight, Power, and Cost) constraints.
Traditional approaches require separate hardware and signal-processing chains for radar and radio, which is bulky and power-hungry. While Integrated Sensing and Communications (ISAC) aims to unify these, existing methods like OFDM-ISAC suffer from high power consumption (PAPR) and lack physical-layer security. Chaos-based waveforms are naturally secure and jam-resistant, but they are notoriously difficult to decode "blindly" (without knowing the transmitter's internal state) in the presence of the high Doppler shifts and multipath interference common in aerial environments.
Methods
The authors propose Ultra-CAN, a Multi-Task Learning (MTL) architecture designed for edge AI.
The Waveform
The system uses Chaos Shift Keying (CSK), where bits are mapped to different chaotic maps (Logistic and Bernoulli). These signals look like noise to an interceptor but have sharp autocorrelation properties ideal for ranging.
The Architecture
Ultra-CAN uses a shared "backbone" to extract features from received I/Q signals. Two variants were tested:
- Conv1D: A two-layer convolutional network with attention pooling.
- QKV Attention: A multi-head Query-Key-Value attention network for capturing long-range temporal dependencies.
The shared representation feeds two specialized "heads":
- Communication Head: A binary classifier that decodes the telemetry bit blindly.
- Sensing Head: A classifier that identifies the specific time-delay (lag) of an obstacle echo.
Key Innovations
- Local-Seed Ranging: In radar mode, the drone uses its own known "seed" to regenerate the chaotic reference locally, allowing it to range its own echoes without any over-the-air state exchange.
- Peer-Exclusion Gate: A safety mechanism that uses the known positions of swarm-mates to "mute" their signals, preventing the radar from mistaking a friendly drone for an obstacle.
- Edge Deployment: The models are quantized to int8 and deployed using TensorFlow Lite Micro on an ESP32-S3 microcontroller.
Datasets
The researchers evaluated the system using synthetic datasets generated from a sophisticated aerial multi-echo channel model.
- Training: 105,000 symbols per scenario, trained across a uniform SNR grid (-5 to 21 dB).
- Evaluation Scenarios: Four distinct aerial swarm environments, including single-echo, multi-echo, and "peer-laden" swarms.
- Out-of-Distribution (OOD) Testing: The models were tested against 3GPP TDL (Tapped Delay Line) channel profiles and various jamming archetypes (barrage, sweep, and reactive followers) to ensure real-world robustness.
Key Insights and Ideas
- Supervision over Capacity: The sensing head outperforms traditional matched filters (66.8% vs 42.4% accuracy) not because it has more parameters, but because it is trained to recognize the geometric "signature" of an obstacle versus clutter.
- The Power of Shared Features: The paper proves that a single 64-dimensional representation is sufficient for both decoding a bit and identifying a radar lag, significantly reducing the computational footprint.
- Jamming Resilience: By combining chaos-based waveforms with frequency agility and "jamming-aware" training, the receiver can maintain a link even under heavy localized interference.
- Quantization Robustness: The transition from floating-point to int8 quantization resulted in negligible performance loss, proving the architecture is well-suited for "bare-metal" hardware.
Key Results
- Accuracy: Achieved a pooled Bit Error Rate (BER) as low as $4.6 \times 10^{-4}$ on single-echo links.
- Sensing Performance: The sensing head identified true obstacles in 66.8% of bursts in crowded swarms, significantly beating the 48% "peak-picker" ceiling of traditional methods.
- Hardware Efficiency: On a 240 MHz ESP32-S3, the Conv1D variant processes a burst in just 16.9 ms, using only 24.2 KiB of memory.
- Energy: Measured energy consumption is a mere 2.2–3.4 mJ per burst, enabling sustained telemetry or a 7.4 Hz range track on a tiny power budget.
Limitations and Future Work
- Simulation vs. Reality: While the models were deployed on physical microcontrollers, the wireless channel was emulated. Real-world over-the-air (OTA) testing is the necessary next step.
- Self-Interference: The radar mode assumes a 40 dB self-interference cancellation (SIC) budget. While realistic for modern full-duplex radios, the actual performance would depend on the specific RF front-end hardware.
- Scalability: Future work aims to explore M-ary chaos-based modulation to increase data throughput beyond the current 59 bit/s limit and investigate cooperative-bistatic sensing where drones help each other range obstacles.
For more details, the authors have released their channel model, data generator, and receiver implementations on GitHub.
Summary updated