source UZH Robotics Perception Group
This video presents a self-adaptive control framework for quadrotors that enables them to learn and adapt in real-world flight scenarios, achieving significant performance improvements in agility and speed in just 100 seconds. The method allows quadrotors to track trajectories accurately, even when faced with changes in payload and environmental conditions, resulting in a threefold increase in flight speed while maintaining safety bounds.
Real-world adaptability
The framework allows quadrotors to learn and quickly adapt their flight policies directly in the physical world, similar to human motor learning. This adaptation enables agile maneuvers such as rapid trajectory tracking, overcoming the limitations of relying on pre-training in simulations.
Self-adaptive framework performance
Using a single control policy, the quadrotor can evolve from conservative maneuvers to maximum actuation limits, tripling its flight speed in approximately 100 seconds, while still maintaining accuracy in trajectory tracking.
Robustness to changes
The method demonstrates robustness when the quadrotor experiences changes in physical characteristics, like additional payloads, quickly adapting to new dynamics and reducing tracking errors effectively, achieving significant speed increases.
Success under disturbances
In tests involving wind disturbances, the adaptive framework allows the quadrotor to efficiently adjust to unexpected challenges, completing tasks with a 50% reduction in travel time compared to its initial conservative policy.
Real-world applications
The self-adaptive control method is shown to be practical for real-world applications such as autonomous inspections, showcasing its effectiveness for complex flight tasks amidst various environmental challenges.
