Emanuel Todorov
homes.cs.washington.edu/~todorov
Projects:
Highly Biomimetic Anthropomorphic Robotic Hand
MuJoCo
Printable View
Emanuel Todorov
homes.cs.washington.edu/~todorov
Projects:
Highly Biomimetic Anthropomorphic Robotic Hand
MuJoCo
https://youtu.be/nXGoU44bqQ4
Dynamic Walking 2013 : Day 4 : Emo Todorov : Dynamic optimization of behavior
Published on Jul 23, 2013
Quote:
June 13, 2013
Dynamic Optimization of Behavior
Emo Todorov, University of Washington
The control of rich dynamic behaviors involving underactuation and contact dynamics remains a mystery. The brain somehow does it, in a way that appears related to optimal control, however the algorithmic details are hard to infer from experimental data. While optimal control is a mature mathematical framework, available algorithms run into the curse of dimensionality. Trajectory optimization has emerged as the most scalable approach. The general idea dates back to the 1950's. Yet until recently, this form of optimization was limited to cleaning up pre-existing trajectories, or to optimizing the smooth segments between manually-specified contact events. And even this these limitations, it was too slow for online re-planning of complex movements.
In this talk I will describe our efforts to make trajectory optimization fully autonomous so that no montion capture or manual scripting is required, and speed it up so that it can be used for online re-planning. Our latest methods have been able to solve every problem we have thrown at them -- including walking, running, getting up, manipulating objects, performing cooperative actions. On a single computer, optimizing a movement without informative initialization takes minutes. Once a solution is found, re-optimizing it online (i.e. warm-starting from the previous iteration) takes tens of milliseconds. While there is still plenty of room for improvement, in particular by using more parallel computing, the existing methods can already solve complex control problems fully autonomously. Unfortunately there is no simple trick to make this happen; instead we have had to work hard on multiple fronts: developing a fast physics engine (MuJoCo), new models of contact dynamics that are realistic yet suitable for differentiation and optimization, and a long list of algorithmic refinements than enable our optimizers to navigate challenging cost landscapes.
Bio
Emo Todorov obtained his PhD in Cognitive Neuroscience from MIT in 1998. He was then a postdoc in Computational Motor Control at University College London, research scientist in Biomedical Engineering at USC, Assistant Professor in Cognitive Science at UCSD, and is now Associate Professor in Applied Mathematics and Computer Science & Engineering at UW. His research focuses on intelligent control in biology and engineering -- in particular using numerical optimization to understand how brains, and some day robots, can autonomously generate complex yet successful movements.
https://youtu.be/jMIKHhZjYPo
Emo Todorov: Synthesis of contact-rich behaviors with optimal control
Published on Jul 26, 2015
Quote:
Brainy Days in Jerusalem:
An interdisciplinary celebration
June 22-25, 2015, Mishkenot Sha’ananim, Jerusalem, Israel
Animals and machines interact with their environment mainly through physical contact. Yet the discontinuous nature of contact dynamics complicates planning and control, especially when combined with uncertainty. We have recently made progress in terms of optimizing complex trajectories that involve many contact events. These events do not need to be specified in advance, but instead are discovered fully automatically. Key to our success is the development of new models of contact dynamics, which enable continuation methods that in turn help the optimizer avoid a combinatorial search over contact configurations. We can presently synthesize humanoid trajectories in tasks such as getting up from the floor, walking and running, turning, riding a unicycle, as well as a variety of dexterous hand manipulation tasks. When augmented with warm-starts in the context of model predictive control, our optimizers can run in real-time and be used as approximately-optimal feedback controllers. Some of these controllers have already been transferred to physical robots, via ensemble optimization methods that increase robustness to modeling errors. The resulting trajectory libraries are also used to train recurrent neural networks. After training the networks can control the body autonomously, without further help from the trajectory optimizer.