Modeling and Control of Off-road Autonomous Vehicles with Situationally Aware Data-Driven Framework (Results Documentation)
This repository serves as a supplemental documentation for capturing the framework, experimental setup and results for Static and Adaptive Multi-Model Parameterized Koopman framework.
In recent years, the scope of deployments of Uncrewed Ground Vehicles (UGVs) has significantly expanded from traditional onroad- and industrial- environments to in- creasingly challenging off-road terrain. These challenging do- mains feature non-linear wheel-terrain interactions and sig- nificant environmental perturbations. This complicates use of traditional model-based control paradigms - accurate models can be both challenging to derive and may not fully capture real-world complexities. In this regard, learning-based tech- niques leveraging neural-networks are often used to capture complex dynamics but often at the cost of explainability and generalizability. This study presents an alternate viewpoint to capture the complex vehicle dynamics in unstructured environments through a data-driven learning-based approach but without the mentioned sacrifices. This is achieved by a novel End-2-End framework utilizing the Koopman operator theory. Our framework captures the dynamics in a multi-model setting with an auxiliary control matrix to factor in environmental perturbations. Additionally, we propose an outer control loop incorporating a novel model-driven kinodynamic motion plan- ner and path-tracking controller. Our approach is computation- ally efficient, robust and generalizable for effectively realizing offroad path tracking problems. This work advances the data- driven modeling- planning- and controls approach for offroad operations, laying the groundwork enabling next-generation UGVs to expand their operating domains even further.
Paper (Preprint):
Modeling and Control of Off-road Autonomous Vehicles with Situationally Aware Data-Driven Framework
- Figures: This folder contains all the plots from the experiment.
- Videos: Folder for storing video results from MMPK hardware deployment.
The framework can be broken down into following process flows.
-
Data-gathering/Model-training (Offline):
- First, we collect data of a vehicle undergoing maneuvers resulting in roll/yaw plane excitation.
- The data is then converted into body-frame representation.
- Finally, the data is parameterized into discrete bins according to curvature.
-
Parameterized family of Koopman models (Offline):
- In a nutshell, the MMPK approach generates model for each of the discrete curvature bin to capture dynamics across the entire operating domain.
- We present two types of modeling techniques based on MMPK approach (more info on fundamental foundation can be found here).
- Static MMPK: Modeling without consideration of effects of offroad terrain leading to load-transfers.
- Adaptive MMPK: Considers and rejects terrain disturbances with the help of an augmented control matrix.
-
Outer control loop (Online):
- For the online real-time deployment, the reference trajectory and the pose estimate of the vehicle are sent to the local motion planner.
- The local planner samples all the feasible trajectories across models and chooses the one closest to the reference trajectory.
- Similar to modeling techniques, there are two novel kinodynamic motion planners presented here.
- Curvature-based Reachability Planner: For each of the model, this planner generates a trajectory satisfying the respective curvature constraint.
- Load-transfer based Reachability Planner: For each of the model, this planner not only generates a trajectory to satisfy the curvature constraint, but also considers the dynamic load transfers due to terrain perturbations to predict pose evolution across the prediction horizon.
- This reference trajectory along with the selected curvature is passed forward to the linear MPC.
- The linear MPC produces high-level commands (velocity steering) for the vehicle.
The following diagram presents pictorial representation for host of experimental stipulations for thorough validation of MMPK framework in simulation and hardware deployment settings.
The experiments are divided into two main sections: Experiment I for simulation-based deployment and Experiment II for hardware deployment, as illustrated above. Each experiment considers a modeling approach (Static or Adaptive MMPK), planning strategy (Curvature-based or Load-Transfer based Reachability Planner), along with a linear MPC considering the chosen model set. The experimental setup includes the following stipulations:
Experiment I-A/II-A:
Static MMPK with Curvature-based Reachability Planner.
Here the modeling, planning and control do not account for adapting to environmental perturbations.
Experiment I-B/II-B:
Adaptive MMPK with Curvature-based Reachability Planner. In this case, the model and subsequent controller adapt to changing terrain. However, similar to Experiment (I-A, I-B), the motion planner simply generates trajectories to satisfy curvature requirements.
Experiment I-C/II-C:
Adaptive MMPK with Load-transfer based Reachability Planner. Finally, all the modalities consider the effect of terrain perturbations. The modeling and control is adaptive as well as the motion planner, which generates trajectories based on dynamically evolving operating conditions.
Simulation deployment (Experiment I-A to I-C) Box plot:
Simulation deployment (Experiment I-A to I-C) CG distribution:
Hardware deployment (Experiment II-A to II-D) Box plot:
Hardware deployment (Experiment II-A to II-D) Tracking Performance:
Hardware deployment videos
 |
 |
|---|---|
| Exp II-A Run 1 | Exp II-A Run 2 |
 |
 |
| Exp II-A Run 3 | Exp II-A Run 4 |
 |
 |
|---|---|
| Exp II-B Run 1 | Exp II-B Run 2 |
 |
 |
| Exp II-B Run 3 | Exp II-B Run 4 |
 |
 |
|---|---|
| Exp II-C Run 1 | Exp II-C Run 2 |
 |
 |
| Exp II-C Run 3 | Exp II-C Run 4 |
Ajinkya Joglekar
This work represents the culmination of two years of research on Koopman Theory applied to Uncrewed Ground Vehicles (UGVs). If you find this research valuable, I encourage you to explore my other papers. If you utilize any part of this dataset in your own research, please consider citing the following papers:
Data-Driven Modeling and Experimental Validation of Autonomous Vehicles Using Koopman Operator
@INPROCEEDINGS{10341797,
author={Joglekar, Ajinkya and Sutavani, Sarang and Samak, Chinmay and Samak, Tanmay and Kosaraju, Krishna Chaitanya and Smereka, Jonathon and Gorsich, David and Vaidya, Umesh and Krovi, Venkat},
booktitle={2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
title={Data-Driven Modeling and Experimental Validation of Autonomous Vehicles Using Koopman Operator},
year={2023},
volume={},
number={},
pages={9442-9447},
keywords={Location awareness;Analytical models;Trajectory tracking;Computational modeling;Predictive models;Vehicle dynamics;Standards},
doi={10.1109/IROS55552.2023.10341797}
}@article{joglekar2024expanding,
title={Expanding Autonomous Ground Vehicle Navigation Capabilities through a Multi-Model Parameterized Koopman Framework},
author={Joglekar, Ajinkya and Samak, Chinmay and Samak, Tanmay and Krovi, Venkat and Vaidya, Umesh},
year={2024}
}@article{joglekar2023analytical,
title={Analytical Construction of Koopman EDMD Candidate Functions for Optimal Control of Ackermann-Steered Vehicles},
author={Joglekar, Ajinkya and Samak, Chinmay and Samak, Tanmay and Kosaraju, Krishna Chaitanya and Smereka, Jonathon and Brudnak, Mark and Gorsich, David and Krovi, Venkat and Vaidya, Umesh},
journal={IFAC-PapersOnLine},
volume={56},
number={3},
pages={619--624},
year={2023},
publisher={Elsevier}
}