PiDR: Physics-Informed Inertial Dead Reckoning for Autonomous Platforms

Researchers propose PiDR, a physics-informed neural network framework for autonomous navigation using only inertial sensors, achieving 29% positioning improvement over conventional approaches. The system addresses critical limitations of traditional deep learning by embedding physical principles directly into the model, enabling accurate dead reckoning in GPS-denied environments without requiring extensive training data.

PiDR represents a meaningful advance in autonomous navigation technology by solving a fundamental challenge: maintaining accurate positioning when external reference signals like GPS are unavailable. This capability is critical for autonomous underwater vehicles, underground robots, and other platforms operating in signal-denied environments where pure inertial measurement is the only viable option. The research bridges the gap between physics-based navigation models and modern machine learning by explicitly embedding inertial navigation equations into the neural network architecture, creating a hybrid approach that maintains physical interpretability while improving accuracy.

Conventional deep learning approaches to this problem suffer from opacity and data inefficiency—they require massive labeled datasets and often violate fundamental physical constraints, leading to implausible trajectory predictions. PiDR's physics-informed design mitigates these issues by constraining the model to respect real-world inertial principles during training, enabling effective learning even with sparse supervision. This matters because real-world sensor data collection is expensive and time-consuming, particularly for specialized platforms like autonomous underwater vehicles.

The practical implications extend across robotics, autonomous systems, and defense applications. Platforms with access to only inertial measurement units—accelerometers, gyroscopes, and magnetometers—can now achieve significantly better positioning accuracy. The 29% improvement demonstrated on diverse platforms suggests strong generalization capability. Additionally, PiDR's lightweight architecture enables deployment on resource-constrained systems, avoiding the computational overhead that would typically accompany advanced neural networks.

Future applications likely include integration with hybrid navigation systems that switch between GPS-denied and GPS-available modes, as well as multi-sensor fusion frameworks where inertial data becomes more reliable between intermittent external updates.

• →Physics-informed neural networks achieve 29% positioning improvement in pure inertial navigation tasks over conventional approaches.
• →The framework enables accurate autonomous navigation in GPS-denied environments using only inertial sensor data.
• →Lightweight architecture allows real-time deployment on resource-constrained platforms like mobile robots and underwater vehicles.
• →Physics-embedded training preserves interpretability and reduces data requirements compared to black-box deep learning models.
• →The approach generalizes across different autonomous platforms and operational environments without platform-specific retraining.