AOI

Accomplishments

• Prepared and gave a NCSRR visiting scholar kickoff talk.
• Met with Jen, Ajay, and Jimmy about 5 week goals.
• Completed most of the orientation here at Stanford.
• Setup a new opensim dev env on linux based on the new git repos.
• Emailed with Jack Wang about modifying the planar walker code and he agreed to let me build of his current work.
• I was introduced to the Wang et. al walker code base by Chris. The code compiles and runs on my machine and I got a copy of the AWS ami to use with my account.
• Read some papers on homotopy and discussed it along with direct collocation with Tom.
• Met with Soha to discuss her work.
• Read some papers to get an idea of scale limits in NLP problems with IPOPT.
• Built most of the rigid body portion of my unactuated planar gait model in Opensim.
• Completed the code for the inverted pendulum on a cart direct collocation problem and tested with IPOPT...but still tracking bugs.
• Many email discussions with Ton back in Cleveland which included a basic parameter id with direct collocation based on one of Tom's homotopy example problems that showed DC can work for parameter id with 5000 nodes (solves in 23s on desktop machine but sensitive to initial guess in parameters). See https://gist.github.com/moorepants/3f122bbad42ca7e910f8 for code.
• Mentored the SymPy/PyDy google summer of code students on various issues: vector calculus, code generation, and visualization.
• Worked on revising my research statement for a prof application.
• Worked on a lingering software paper with Chris to finalize and send it to F1000Research.

Objectives

• Read and understand Jack and Tim's code focusing on the objective function evaluation. Make a plan to implement the tracking objective.
• Sniff out the errors in the inverted pendulum dc problem and get it running for a large number of nodes. Implement an implicit backward Euler integrator for the simulation to match the DC discretization I'm using.
• Complete the 2D gait opensim model and compare it's output to the custom models I have.
• Design a gain scheduled full state feedback controller in opensim for the 2D gait model that matches the one used in the direct approach system id method.

Issues

• I still have a bug preventing the inverted pendulum direct collocation problem from converging. I think I've narrowed it down to the differences in integration routine (LSODA: Adams + BDF) for the simulation and the discretization method (backward Euler) for the direct collocation.
• Never got all the paperwork and stuff together to get a building key. Abandoned...
• I'm already a half a week or so behind what I proposed for the 5 week period here.