Dr. Jonathan P. "Jon" How, Professor of Aeronautics and Astronautics at the Massachusetts
Institute of Technology (MIT) was the March speaker in our monthly seminar series on autonomy. He presented recent algorithms and
results for the safe, real-time motion planning under uncertainty of single and
multi-agent systems in dynamic environments. In order to safely navigate in a
dynamic environment, an autonomous system must be able to overcome uncertainty
in both its own motion and the motion of other agents. Robust planning can be
achieved by embedding probabilistic uncertainty models into a
chance-constrained RRT* planner (CC-RRT*). This algorithm leverages the
sampling-based nature of RRTs to generate probabilistically feasible solutions,
with guarantees on maximum risk of constraint violation, in real-time. CC-RRT*
has been demonstrated to efficiently produce risk-aware trajectories for a
variety of complex aerospace-related motion planning problems, including
applications for urban driving and parafoil terminal guidance.
Reliable robust planning also
depends on the availability of probabilistic models that accurately represent
the uncertainty in the environment and its evolution. This talk will thus
present recent results for efficiently learning these models for dynamic
environments. We utilize Bayesian nonparametric models that uniquely provide
the flexibility to learn model size and parameters, which are often very
difficult to determine a priori. For example, Gaussian processes (GPs) are used
to represent the trajectory velocity fields of the obstacles (static &
dynamic) in the environment, a Dirichlet process GP mixture (DP-GP) is used to
learn the number of motion models and their velocity fields, and the dependent
Dirichlet process GP mixture (DDP-GP) is also used to capture the same
quantities and their temporal evolution. These techniques have been used to
learn models of motion/intent behaviors of other drivers and pedestrians to
improve the performance of an autonomous car.
Here are his charts:
