Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Game Theoretic Planning for Autonomous Vehicles
Autonomous vehicles regularly interact with others, e.g., when overtaking, merging, and changing lanes. A promising technique for managing these interactions is game theoretic planning, wherein the behaviour of other vehicles are modelled using game theory.
Keywords: Autonomous Vehicles, Multi-agent Planning, Game Theory, Numerical Methods
Reasoning about the intentions and responses of others is a challenging but essential task in motion planning and control of autonomous vehicles. Interactions typically include both cooperation (i.e., mutual collision aversion) and competition (e.g., overtaking, blocking), and formulating planning and control strategies that capture these rich behaviours is a topic of ongoing research. A promising approach is game theoretic planning (GTP) where the ego agent reasons about the responses of other agents to their actions by computing a Nash equilibrium. In GTP, the ego agent typically computes a Nash equilibrium over a finite horizon, takes an action, then re-plans à la model predictive control (MPC).
Implementing GTP requires computing (approximate) Nash equilibria in real time; this is usually accomplished by solving parameterized variational inequalities (VIs) at each sampling instant. These VIs are invariably non-monotone and are challenging to solve using onboard computing hardware. Nevertheless, there has been progress in recent years using heuristic algorithms such as modified best response or iterative linear-quadratic games as well as the more systematic augmented Lagrangian framework.
In this project, we aim to develop methods for solving real-time GTP problems by drawing on the theory of VIs. We will explore several classes of algorithms for solving VIs such as Josephy-Newton, interior-point, or semi-smooth Newton methods, specialize them to GTP problems, and implement the most promising on the ORCA testbed.
Reasoning about the intentions and responses of others is a challenging but essential task in motion planning and control of autonomous vehicles. Interactions typically include both cooperation (i.e., mutual collision aversion) and competition (e.g., overtaking, blocking), and formulating planning and control strategies that capture these rich behaviours is a topic of ongoing research. A promising approach is game theoretic planning (GTP) where the ego agent reasons about the responses of other agents to their actions by computing a Nash equilibrium. In GTP, the ego agent typically computes a Nash equilibrium over a finite horizon, takes an action, then re-plans à la model predictive control (MPC).
Implementing GTP requires computing (approximate) Nash equilibria in real time; this is usually accomplished by solving parameterized variational inequalities (VIs) at each sampling instant. These VIs are invariably non-monotone and are challenging to solve using onboard computing hardware. Nevertheless, there has been progress in recent years using heuristic algorithms such as modified best response or iterative linear-quadratic games as well as the more systematic augmented Lagrangian framework.
In this project, we aim to develop methods for solving real-time GTP problems by drawing on the theory of VIs. We will explore several classes of algorithms for solving VIs such as Josephy-Newton, interior-point, or semi-smooth Newton methods, specialize them to GTP problems, and implement the most promising on the ORCA testbed.
- Learn about Receding Horizon Planning, Autonomous Driving, and Numerical Methods for Non-monotone Games.
- Mathematically formulate a game theoretic planner for a multi-vehicle autonomous driving scenario.
- Implement and compare interior-point and Newton-based methods for solving non-monotone games against algorithms from literature using the simulated autonomous driving scenario as a case study.
- Implement your algorithm on the ORCA autonomous race cars.
- Learn about Receding Horizon Planning, Autonomous Driving, and Numerical Methods for Non-monotone Games. - Mathematically formulate a game theoretic planner for a multi-vehicle autonomous driving scenario. - Implement and compare interior-point and Newton-based methods for solving non-monotone games against algorithms from literature using the simulated autonomous driving scenario as a case study. - Implement your algorithm on the ORCA autonomous race cars.
- Please send your resume/CV (including lists of relevant publications/projects) and transcript of records in PDF format via email to dliaomc@ethz.ch, gbanjac@ethz.ch, and alex.liniger@vision.ee.ethz.ch.
- We anticipate this project starting in September/October 2021.
- Please send your resume/CV (including lists of relevant publications/projects) and transcript of records in PDF format via email to dliaomc@ethz.ch, gbanjac@ethz.ch, and alex.liniger@vision.ee.ethz.ch. - We anticipate this project starting in September/October 2021.