Motion and Behaviour Planning
Decision Making Under Uncertainty
Motion Planning and Decision Making
Our motion planning, behaviour planning and decision making under uncertainty research focuses on enabling autonomous vehicles to deal with both the adversarial traffic and traffic with highly stochastic dynamics.
We are researching in several areas like motion planning, reinforcement learning, non-linear mathematical optimization, probabilistic graphical models, apprenticeship learning, inverse reinforcement learning, game theory, heuristic search, and graph theory to develop novel algorithmic frameworks to deal with complex traffic configurations with highly stochastic and adversarial dynamics.
We research in maching learning, deep learning and computer vision to create algorithms for perceiving both structured and unstructured environments using cameras. Our deep learning research for perception focuses on coming up with better network architectures, constraints and formulations, cost functions etc. to ensure that the our deep neural networks (a) are data-efficient, (b) generalize well, and (c) have real-time inference capabilities.
Currently our research is headed towards enabling autonomous vehicles to perceive both during day and night using off-the-shelf RGB and NIR cameras.
Redundancy of High-Fideliy Maps
Our perception and behaviour planning algorithms make high-definition (high-fidelity) 3D-mapping of environments redundant. We do not require dense maps for automous driving, and can use just GPS maps for end-to-end navigation. Our autonomous driving technology enables vehicles to localize with respect to the local delimiters in the environment, and use GPS maps for high-level route planning.
We provide 4 different kinds of mapping for different applications: (a) high-fidelity mapping, (b) sparse mapping, (c) GPS mapping combined with our autonomous driving technology, and (d) dynamic polyhedral sets based maps (DPS) which is our proprietary mapping solution.
Technology Demonstration Videos