PIAug -- Physics Informed Augmentation for Learning Vehicle Dynamics for Off-Road Navigation

Parv Maheshwari, Wenshan Wang, Samuel Triest, Matthew Sivaprakasam, Shubhra Aich, John G. Rogers, Jason M. Gregory, Sebastian Scherer

Off-road navigation is a "wild west" for robotics. Unlike the predictable asphalt of city streets, off-road terrain involves unpredictable tire slip, varying inclines, and complex vehicle dynamics. Traditionally, researchers have relied on either rigid physics models that break down at high speeds or data-driven neural networks that fail the moment they encounter a scenario they haven't seen before.

A new paper, PIAug, introduces a clever middle ground: using simple physics to "hallucinate" high-speed data from low-speed observations, allowing robots to handle aggressive maneuvers they were never explicitly trained for.

TL;DR

PIAug is a data augmentation framework that uses a nominal physics model to generate synthetic high-velocity training data from real-world low-velocity datasets. By incorporating a physics-inspired loss function, the resulting neural network achieves up to 67% less error in trajectory prediction and enables 4x tighter waypoint tracking in real-world aggressive driving scenarios.

What Problem Does This Paper Address?

Modeling vehicle dynamics off-road is a balancing act.

  1. Physics-based models (like the Kinematic Bicycle Model or KBM) are great for generalization but rely on the "no-slip" assumption. At high speeds or on loose dirt, this assumption fails, leading to massive prediction errors.
  2. Data-driven models (Neural Networks) can learn complex, non-linear dynamics from raw data. However, they are notoriously "brittle." If you train a robot to drive at 2 m/s, it will likely fail or behave unpredictably when pushed to 5 m/s.

The core challenge is that collecting high-speed, aggressive off-road data is dangerous and expensive. We need a way to train models for high-speed "target domains" using only safe, low-speed "source domain" data.

Methods

The authors propose PIAug (Physics Informed Augmentation). The workflow happens during the training phase:

  • State Scaling: The system takes a real-world "mini-batch" of low-speed data. It then applies a random scaling factor to the initial velocities, effectively "speeding up" the starting state of the robot.
  • Physics-Based Labeling: Since we don't have real ground truth for what happens next at that high speed, the system uses the Extended Kinematic Bicycle Model (KBM) to project a multi-step future trajectory.
  • Hybrid Loss Function: The Neural Network (NN) is trained using two losses simultaneously:
  • Data Loss ($L_D$): Standard supervised learning on the real, low-speed data.
  • Physics Loss ($L_P$): A loss that penalizes the NN if its predictions for the augmented (high-speed) data deviate from what the physics model (KBM) predicts.
  • Multi-modal Inputs: The NN doesn't just look at numbers; it processes FPV images, top-down RGB maps, and 4-channel height maps (min, max, mean, std) to understand the terrain context.

Datasets

  • TartanDrive: The authors used a subset of this large-scale off-road dataset, specifically filtering for low-velocity trajectories ($<3$ m/s) to act as the training source.
  • Balanced Velocity Dataset: A new dataset gathered by the authors containing a wider distribution of speeds to evaluate how well the model generalizes to out-of-distribution (high-speed) scenarios.

Key Insights and Ideas

  • Physics as a Prior, Not a Ceiling: The KBM is imperfect, but it provides a "kino-dynamically feasible" baseline. By using it to generate augmented data, the NN learns the structure of high-speed motion even if the KBM's specific predictions aren't 100% accurate to real-world slip.
  • Continuous Rotation Representation: Instead of using quaternions (which have mathematical discontinuities), the authors used a 6-state continuous rotation representation. This makes it easier for the neural network to learn smooth transitions in orientation.
  • Computational Efficiency: Interestingly, the authors found that an optimized NN (using GPU-based RNNs) can actually perform inference faster than solving the differential equations for a KBM during real-time planning (MPPI).

Key Results

  • Accuracy: In high-speed aggressive turns, PIAug reduced mean position error by up to 67% compared to a standalone KBM.
  • Generalization: PIAug significantly outperformed "Vanilla" NNs and standard Physics-Informed Neural Networks (PINNs) when tested on speeds 2-3x higher than the training data.
  • Real-World Success: In field tests using a modified Yamaha Viking ATV, the PIAug-trained model successfully navigated a figure-8 track with a 1m goal radius at high speeds—a constraint 4x tighter than what the KBM could handle before failing.

Limitations and Future Work

While PIAug bridges the velocity gap, it doesn't yet solve the environment gap. A model trained on forest data might still struggle in an open desert because the visual features (FPV/RGB maps) change drastically.

The authors suggest three future directions:

  1. Robot Transfer: Using PIAug to quickly adapt a model from one robot platform to another using the target robot's nominal model.
  2. Complex Physics: Integrating more sophisticated models (like the "Magic Formula" for tire forces) to further narrow the gap between simulation and reality.
  3. Constraint-Informed Sampling: Instead of random scaling, sampling augmentations based on specific safety or task constraints.

Summary updated

Abstract

Modeling the precise dynamics of off-road vehicles is a complex yet essential task due to the challenging terrain they encounter and the need for optimal performance and safety. Recently, there has been a focus on integrating nominal physics-based models alongside data-driven neural networks using Physics Informed Neural Networks. These approaches often assume the availability of a well-distributed dataset; however, this assumption may not hold due to regions in the physical distribution that are hard to collect, such as high-speed motions and rare terrains. Therefore, we introduce a physics-informed data augmentation methodology called PIAug. We show an example use case of the same by modeling high-speed and aggressive motion predictions, given a dataset with only low-speed data. During the training phase, we leverage the nominal model for generating target domain (medium and high velocity) data using the available source data (low velocity). Subsequently, we employ a physics-inspired loss function with this augmented dataset to incorporate prior knowledge of physics into the neural network. Our methodology results in up to 67% less mean error in trajectory prediction in comparison to a standalone nominal model, especially during aggressive maneuvers at speeds outside the training domain. In real-life navigation experiments, our model succeeds in 4x tighter waypoint tracking constraints than the Kinematic Bicycle Model (KBM) at out-of-domain velocities.
BibTeX
@misc{maheshwari2023piaug,
  author = {Parv Maheshwari and Wenshan Wang and Samuel Triest and Matthew Sivaprakasam and Shubhra Aich and John G. Rogers and Jason M. Gregory and Sebastian Scherer},
  title = {{PIAug -- Physics Informed Augmentation for Learning Vehicle Dynamics for Off-Road Navigation}},
  year = {2023},
  eprint = {2311.00815},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2311.00815},
}
Figure