AUV Acceleration Prediction Using DVL and Deep Learning

Autonomous underwater vehicles (AUVs) are essential for various applications, including oceanographic surveys, underwater mapping, and infrastructure inspections. Accurate and robust navigation are critical to completing these tasks. To this end, a Doppler velocity log (DVL) and inertial sensors are fused together. Recently, a model-based approach demonstrated the ability to extract the vehicle acceleration vector from DVL velocity measurements. Motivated by this advancement, in this paper we present an end-to-end deep learning approach to estimate the AUV acceleration vector based on past DVL velocity measurements. Based on recorded data from sea experiments, we demonstrate that the proposed method improves acceleration vector estimation by more than 65% compared to the model-based approach by using data-driven techniques. As a result of our data-driven approach, we can enhance navigation accuracy and reliability in AUV applications, contributing to more efficient and effective underwater missions through improved accuracy and reliability.

💡 Research Summary

The paper addresses a critical need in autonomous underwater vehicle (AUV) navigation: accurate estimation of the vehicle’s acceleration vector using only past Doppler Velocity Log (DVL) velocity measurements. Traditional navigation systems fuse inertial measurement unit (IMU) data with DVL velocity through extended Kalman filters, but the inertial component suffers from drift, and the DVL‑only model‑based approach recently proposed by Klein and Lipman computes acceleration via a simple least‑squares formulation that assumes a linear relationship between velocity and elapsed time. While mathematically straightforward, this method cannot capture the complex, nonlinear dynamics introduced by hydrodynamic forces, varying currents, and sensor noise that are typical in real underwater environments.

To overcome these limitations, the authors propose an end‑to‑end deep learning (DL) architecture that learns the mapping from a short history of DVL velocity vectors to the corresponding three‑axis acceleration vector. The network combines a one‑dimensional convolutional neural network (CNN) with two stacked long short‑term memory (LSTM) layers. The CNN (kernel size 3, LeakyReLU activation with α = 0.01) extracts localized temporal features such as abrupt velocity changes, while a max‑pooling layer (pool size 3) reduces dimensionality and highlights salient patterns. The LSTM layers (32 hidden units each) capture longer‑range dependencies, allowing the model to infer how past velocity trends influence current acceleration. Two fully‑connected layers with LeakyReLU produce the final acceleration estimate. Training uses mean‑squared error loss, a batch size of 16, a learning rate of 0.001, and runs for 30 epochs.

The experimental dataset originates from sea trials of the Snapir AUV in the Mediterranean Sea near Haifa, Israel. Snapir is equipped with a high‑performance fiber‑optic gyroscope (FOG) IMU operating at 100 Hz and an RDI Workhorse DVL sampling at 1 Hz, delivering velocity measurements with a standard deviation of 0.02 m/s. From the full mission, 66.7 minutes of data (10 segments, each 400 s) were extracted for training. To improve robustness, the authors performed data augmentation by adding artificial scale, bias, and Gaussian white noise, effectively doubling the training set to 133.4 minutes (20 segments). Three additional 400‑second segments, completely held out from training, served as the test set, ensuring an unbiased evaluation. Ground‑truth acceleration for training and testing was derived from the INS/DVL fused solution, providing a reliable reference.

Performance is assessed using root‑mean‑square error (RMSE). The proposed DL model achieves an RMSE of 0.082 m/s² on the test set, compared to 0.25 m/s² for the baseline least‑squares model—a 67.2 % reduction in error. Training and validation loss curves show smooth convergence within the 30‑epoch window, indicating stable learning without over‑fitting. The authors highlight that the DL approach consistently improves accuracy across varying depths, speeds, and maneuvering patterns, demonstrating adaptability to diverse operational conditions.

In the discussion, the authors argue that the learned nonlinear mapping enables the model to compensate for hydrodynamic effects and sensor imperfections that the linear baseline cannot address. They note that the method requires only software changes, making it readily deployable on existing AUV platforms without hardware modifications. Potential benefits include the ability to use lower‑grade sensors while maintaining high navigation performance, which could lower costs for small or micro‑AUVs.

The paper concludes by emphasizing that integrating the DL‑derived acceleration estimates into the INS/DVL fusion loop is expected to further improve overall navigation accuracy and convergence speed, as suggested by prior work that incorporated DVL‑based acceleration. Future research directions include expanding the dataset to multiple vehicle types and oceanic regions, exploring model compression for real‑time onboard execution, and quantitatively evaluating the impact of the estimated acceleration on full INS/DVL state estimation.

Overall, the study demonstrates that a relatively simple CNN‑LSTM architecture can substantially outperform a physics‑based least‑squares estimator for AUV acceleration, offering a practical pathway to more reliable and precise underwater navigation.