Developing novel controls for robotic systems operating safely in complex environments
Researchers in APL’s Intelligent Systems Center (ISC) create capabilities that give robots the speed, agility, intelligence, intuition, physical characteristics, and resiliency required for both safe operation in the real world and effective teaming with human partners. Research focus areas include:
To test the NMPC algorithm, ISC researchers explored the receding-horizon case of navigating through a narrow corridor. They evaluated the algorithm in both simulation and hardware on a 24-inch-wingspan UAV, demonstrating agile fixed-wing flight using offboard sensing and computation.
This approach was also tested using onboard sensing and computation on a 42-inch-wingspan UAV operating in an urban environment.
APL researchers demonstrated the improved performance in post-stall maneuver planning enabled by unsteady aerodynamics within a UAV’s flight dynamics model. A vortex particle model of local aerodynamics is integrated into model-based control for a morphing-wing UAV. This UAV’s performance is compared for gliding perch maneuvers with and without wing morphing and unsteady aerodynamics. Wing morphing planned using quasi-steady aerodynamics dramatically reduced performance, but wing morphing planned with unsteady aerodynamics improved performance. However, this result required pre-planning time; unsteady solutions generated in flight did not converge quickly enough for improved control.
The second domain ISC researchers explored was precision post-stall landing. Fixed-wing UAVs typically require long, flat runways. The objective of this effort was to land precisely with minimum speed and in minimum distance. Researchers used as little thrust as possible to land from a significant cruise speed and altitude (above the tree line). They explored both a trajectory library approach as well the aforementioned NMPC approach. Both algorithms leveraged radial basis function neural networks to more precisely model the lift, drag, and moment coefficients associated with the post-stall regime. The algorithm was tested on a 60-inch-wingspan, 4.5-kg fixed-wing UAV and demonstrated precision landing in variable wind.
Safe high-speed navigation is a key enabling capability for real-world deployment of robotic systems. A significant limitation of existing approaches is the computational bottleneck associated with explicit mapping and the limited FOV of existing sensor technologies. Researchers in the ISC study algorithmic approaches that allow robots to predict spaces extending beyond the sensor horizon, opening the possibility of robust planning at high speeds and more efficient map exploration.
They accomplish this by using a generative neural network trained from real-world data. They extend their existing control algorithms to support leveraging the predicted spaces to improve collision-free planning and navigation at high speeds. Using a physical robot based on the MIT RACECAR equipped with an RGBD sensor, a team from the ISC was able to show improved navigation performance at high speeds, as well as increased efficiency in exploring and mapping new areas.
ISC researchers integrated the NanoMap framework with their NMPC approach to enable vision-based collision-free navigation with fixed-wing UAVs. As a result, they reached cruise speeds of 20 mph and demonstrated the ability to execute aggressive maneuvers to avoid obstacles. This approach allows for reasoning about sensing uncertainty, both in the state estimates as well as in the depth measurements used for navigation.
Off-road environments contain a wide range of terrains that pose challenges for conventional autonomous navigation systems. In particular, deformable obstacles, such as grasses, bushes, and other foliage, can be difficult to navigate because these objects, although traversable, appear untraversable when viewed with depth sensors.
To address this challenge, a team from the ISC developed a neural network architecture that uses camera imagery to learn a traversability metric. They collect training data by operating an uncrewed ground vehicle in an environment and evaluating the trajectory deviation error. Objects that are traversable generate a low amount of error, while untraversable objects produce large errors. Given a sampled vehicle trajectory and a history of images, the neural network can learn to predict model error. Using this predictive network, researchers are able to traverse long distances through grasslands, shrubs, and bushes and demonstrate robustness to environments outside the training distribution.
