Creating a safe, dependable, and robust control methodology for AGV autonomous navigation in unknown cluttered environments has long been recognized as a significant challenge. Such a navigation task requires the AGV to navigate safely and autonomously, to avoid getting trapped or colliding with static and dynamic obstacles while moving towards the goal, and to adhere to various system constraints. To that end, the robot should be capable of perceiving and reacting appropriately to its surroundings. As a result, a complex optimization control problem emerges that is difficult to solve in real-time.
A team from the Vehicle Autonomy and Intelligence Lab (VAIL) at Indiana University – Bloomington is working on exactly such challenges with Jackal UGV, with the goal of achieving collision-free navigation of autonomous ground vehicles (AGVs) in unknown cluttered environments such as dense forests, crowded offices, corridors, and warehouses. VAIL specializes in developing methodologies that improve the autonomy and intelligence of robotic systems such as unmanned ground, aerial, and aquatic vehicles. Their research interests include planning, learning, and coordination techniques for single or multiple robots, with potential applications such as autonomous navigation, surveillance security, search and rescue, environmental monitoring, and smart transportation.
Flexible Collision-Free Navigation Method
Sampling-based Model Predictive Control (MPC) methods, such as Model Predictive Path Integral (MPPI), are becoming increasingly popular for many robotics tasks. However, a limitation of these existing methods is that they sometimes generate infeasible paths. To address this issue and to improve MPPI performance, the VAIL team proposed a new strategy, called log-MPPI, which provides more flexible and efficient distributions of sampled trajectories.
Sampling From Normal Log-normal (NLN) Mixture
The main idea behind log-MPPI is that trajectories (or control input updates) are sampled from the normal log-normal distribution (NLN) rather than from the Gaussian distribution. The resulting trajectories are more flexible and more efficient than those generated by the classical MPPI. This results in better exploration of the state spaces and reduces the chance of getting stuck. Further, by storing information about obstacles in the robot’s environment in a 2D occupancy grid map and by incorporating that into the optimization problem, it was possible to have collision-free navigation in unknown cluttered environments.
“It was a great experience to see Jackal UGV navigate safely in cluttered environments not mapped beforehand with its maximum speed, which is 2 m/s, and with full autonomy while avoiding collision with static and dynamic obstacles. Furthermore, it was exciting to see Jackal navigate autonomously in the real world without adjusting the controller parameters, thanks to the Jackal UGV simulated robot, which mimics the real robot.”
– Ihab Mohamed, Ph.D. student
Jackal UGV Supports a Multi-Robot Approach
The VAIL team used three platforms to develop the above methodology: Jackal UGV for indoor and outdoor navigation tasks, Husky UGV for unstructured off-road/terrain and autonomous navigation, and Aquaticus for monitoring their aquatic environments, such as the oceans. The Jackal UGV was outfitted with a 16-beam Velodyne LiDAR sensor, which was used for generating the local costmap and estimating the robot’s pose using the Lidar odometry and mapping (LOAM) ROS package.
This robotic platform proved to be the ideal tool for the lab because it is small, fast, and allows direct control of its movements and rotations. Furthermore, the MPPI algorithm was able to control the Jackal UGV directly via the ROS API with quick ROS sync functionality with an RVIZ GUI and a Gazebo model, making the transition from simulation to real robot relatively simple. Finally, the team quickly adjusted their research due to Jackal UGV’s flexibility in terms of being able to easily plug sensors, cameras, and other accessories into its simple mounting platform.
“It was a great experience to see Jackal UGV navigate safely in cluttered environments not mapped beforehand with its maximum speed, which is 2 m/s, and with full autonomy while avoiding collision with static and dynamic obstacles. Furthermore, it was exciting to see Jackal navigate autonomously in the real world without adjusting the controller parameters, thanks to the Jackal UGV simulated robot, which mimics the real robot,” said Ph.D. student Ihab Mohamed. “We love the dimension of the Jackal UGV, and it has the right size that we need for our navigation projects. Agility is also an important factor that we consider, and the fast-running speed is what we want for evaluating our robotics algorithms in various complex testing scenarios”. said Assistant Professor Lantao Liu.
The VAIL team was overjoyed with their success, as they were able to demonstrate their control strategy for 2D grid-based navigation in unknown cluttered environments such as forest-like and indoor environments in both simulation and real-world settings. To the best of their knowledge, the team believes this is the first experimental attempt at grid-based collision-free navigation using sampling-based MPC, specifically MPPI and log-MPPI. You can read the full paper detailing their work here.
The project has been performed by Ihab S. Mohamed, Dr. Kai Yin, and Dr. Lantao Liu.
To learn more about the Vehicle Autonomy and Intelligence Lab, you can visit their website here.
To learn more about Jackal UGV, you can visit our website here.