Predictive control for belief-space planning
Abstract
In recent years, a significant research effort has been put into autonomous systems. Autonomous systems are systems that are designed to be able to solve a mission without, more or less, any human intervention. Important examples include drones, self-driving cars, Mars rovers, smart software systems.... Common for these systems is that they need to gather information from their surrounding and continuously make decisions that make the system achieve its goals, and this despite uncertainty in the system’s dynamics, sensors, and in the surrounding environment. Examples of uncertain parts of the surrounding environment are other road users (vehicle applications), targets (military applications), and human collaborators (robot applications), whose future motions are in general not known. In simple problems, the uncertainty can be neglected, while for more advanced problems it has to be more explicitly considered which is very computationally demanding. The objective with the proposed research is to advance the computational efficiency of POMDP algorithms for motion planning such that they become more useful to larger problems of higher practical relevance. An important overarching challenge in the proposed project is to develop algorithms that are able to compute safe plans, optimizing important aspects such as energy, time, and risk for robots in dynamic uncertain environments. Furthermore, this should be done sufficiently well while not exceeding the computational resources, also for problems considered large today. Model Predictive Control (MPC) is a class of control algorithms that compute a sequence of control signals that optimizes the predicted behavior of a system. It has its origin in Dynamic Matrix Control (DMC) which was invented by engineers at Shell Oil in the early 1970’s. MPC is originally a tool for control, where open-loop optimal control problems are solved over a finite horizon but only the first control signal is actually applied to the system. In the next time step, the horizon is moved one step ahead and the problem is resolved starting from a newly obtained measurement from the controlled plant. This strategy is called receding horizon control and since new measurements are used in each sample, feedback is introduced. MPC can today be applied to linear systems, hybrid systems and nonlinear systems. The predictive control concept is very general and is expected to be one important tool in the proposed work. The proposed research considers motion planning problems under uncertainty, using non-trivial stochastic motion and measurement models. Planning problems where the state’s probability distribution is explicitly considered are known as belief-space planning problems and can formally be formulated as POMDPs. However, POMDPs are known to be intractable to be solved in its general form. As a result, the exact solution is often out of reach and the challenge becomes to obtain a good trade-off between solution quality and computational tractability. Therefore, the aim with the proposed research is to develop improved algorithms that are able to solve practically relevant theoretically sound approximations to these problems. To achieve this, we plan to extend our recently presented two-step motion planning framework where methods from AI and optimal control are tightly integrated. The main motivation for this two-step approach is that it combines the more combinatorial focus in AI methods with the focus on more advanced continuous dynamicsand optimality aspects in optimal control; first is a state-of-the-art on-line POMDP algorithm with a suitable approximation level applied, then this solution is used to warm-start a numerical optimization local approximation-formulation of the same problem. In particular, the area of application of these state-of-the- art motion planning tools for deterministic problems will be extended to planning problems including stochastic uncertainty.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jan 04, 2023
- Source ID
- FA86552217167
Entities
People
- Daniel Axehill
Organizations
- Air Force Office of Scientific Research
- Linköping University
- United States Air Force