Explore aggressive autonomous flight planning for unmanned aircraft in unknown, cluttered environments through this 59-minute lecture by Nicholas Roy, Associate Professor at MIT. Delve into approximate inference and planning algorithms that enable fast, agile motion for aerial and ground vehicles. Learn about challenges in obstacle detection, rapid decision-making with incomplete data, and innovative approaches to navigation and control. Discover techniques like fast approximation, differential flatness, and RRT* initialization for complex dynamics. Examine strategies for unknown environments, including collision probability learning and generalized kernel estimation. Investigate empirical vs. guaranteed safety, receding-horizon planning, and autonomous RC car experiments. Gain insights from Roy's extensive research in robotics, unmanned aerial vehicles, and artificial intelligence, drawing from his experience as founder of Google's Project Wing.
UW CSE Robotics - Nicholas Roy, "Planning to Fly and Drive Aggressively"
Paul G. Allen School via YouTube
Overview
Syllabus
Intro
Challenges
Fast Approximation Techniques For Planning Aggressive Flight
Planning with Complex Dynamics
The problem with the RRT
Differential flatness
Differentially flat representations exist for many systems
RRT* as a initialization
Performance Comparison
How does the polynomial optimization work?
Unknown Environment Assumptions
Different Cost Functions
Collecting Training Data
Learning Collision Probabilities
Planning With Collision Probabilities
What is the biggest thing holding back the performance?
Generalization
Modelling distributions
Enforcing smoothness
Theorem
Generalized Kernel Estimation
Our solution: A Bayesian approach
Empirical vs Guaranteed Safety
Safety guarantee emphasizes sensing
Receding-horizon planning
Strategy: Predict a correction to the shortest-path heuristic
Training: Brief sketch
Data Representation Descriptive Features
Results: Characteristics of Learned Model
Autonomous RC Car Experiments
Experiment: Baseline Planner
Experiment: Our Planner
What's Next...?
What is Needed?
Acknowledgements
Taught by
Paul G. Allen School
