16-745: Dynamic Optimization: DRC-related Project Suggestions
The project will involve performing a substantial dynamic optimization,
and writing a paper about it. The writeup is as important as the programming
(if not more so) and will be in the format of a conference paper
(more on that later). Those of you who already have a dynamic optimization
problem you are working on for your research should work on that (subject
to the Professor's approval). Another option is to work on the problem
or algorithm
described below (subject
to the Professor's approval).
Optimize DRC Tasks
We will give you models and simulator to work with.
Rough Terrain Walking
Climb Steep Stairs or a Ladder
Getting Into And Out Of A Car
Move Debris
Go Through Doors
Combinatorial Optimization: Efficient Search Over Discrete Actions
Parallel A* For Footstep Planning
How can we use a GPU, multiple threads/cores, or the vector instruction
set of a CPU to make A* more effective?
Choosing Motor Primitives And Their Arguments
How can we choose from a library of discrete behaviors which have
continuous arguments?
Trajectory Optimization
Newton-method-based Trajectory Optimization
Figure out efficient analytic 2nd derivatives for inverse dynamics,
and an optimization algorithm to use this information.
Vector SIMD-based Trajectory Optimization
Maximize performance of trajectory optimization using the vector instruction
set (AVX) of current Intel processors.
GPU-based Trajectory Optimization
Maximize performance of trajectory optimization using a Tesla K20 GPU.
Using Inverse Dynamics-based Trajectory Optimization When No or Multiple
Inverses Exist
How do we handle underactuated systems, double support in walking,
and contact in manipulation?
Designing Robust Behavior
Minimize Risk in Footstep Planning
How can we use probabilistic modeling to choose behavior to keep the
risk of failure below a threshold.
Efficient Policy Optimization Using Parallel Computation
Policy Optimization Using Multiple Models
How can we efficiently optimize a policy over a set of multiple models?
Inverse Optimal Control
Use inverse optimal control to establish what a reasonable optimization
criterion might be for rough terrain walking.
Optimize Perception For DRC Tasks
Treat designing a perception algorithm for doing a DRC task as an
optimization problem, with fixed computational resources and a library
of possible computational operations. A*, trajectory optimization,
and/or policy optimization might work.