Personal tools
You are here: Home CG seminar 2019 Learning to Plan
« August 2019 »
August
SuMoTuWeThFrSa
123
45678910
11121314151617
18192021222324
25262728293031
Log in


Forgot your password?
 

Learning to Plan

Wednesday, March 13th, 2019, 16:10

Schreiber 309

underline

Learning to Plan

Aviv Tamar, The Technion

Abstract:

Motion planning is fundamental to almost all robotic applications deployed today. However, for domains where the environment can change rapidly, such as in e-commerce applications or home robotics, it is desired to plan *fast*, and the computational burden of conventional planning algorithms can be limiting. 
 
Recent developments in deep learning demonstrated learning of complex patterns in data for various decision making domains such as computer vision, protein folding, and games. Motivated by these successes, we ask -- can we use deep learning to approximate a planning computation? In such a scheme, learning offline on a set of training problems would allows us to solve similar problems at test time faster, using the neural network's prediction.
 
We begin by investigating the capacity of deep networks to represent a planning computation. We show that the classic value iteration algorithm is equivalent to a certain type of convolutional network, and exploit this idea to design the value iteration network -- a differentiable planner that can be trained using supervised learning or reinforcement learning. 
Then, we show how deep learning can be used to improve discrete task planning, by learning powerful image-based heuristic functions for A* search. We conclude with recent work on learning for continuous motion planning. Here, the key is in the learning algorithm -- we find that learning to navigate tight passages is fundamentally challenging for supervised learning approaches. We instead propose a novel reinforcement learning algorithm that exploits features of the motion planning problem to effectively learn a neural motion planner. We demonstrate predicting motion plans with almost perfect accuracy on a 4-dimensional robotic arm domain with challenging narrow passages. 

  

Document Actions