We have submitted a new journal paper, PROXDDP: Proximal Constrained Trajectory Optimization.

Check out our preprint here.

Last time we had an introductory look at OT, the definition and computation of the Wasserstein distance.

A few years ago, I wrote a solver for this variational formulation of mean-field games, as part of a project. This eventually led to a small C++ library as a toy project, which might or might not be usable.

Our paper Enforcing the consensus between Trajectory Optimization and Policy Learning for precise robot control (authored by Quentin le Lidec) was accepted at ICRA 2023.

There's been a lot of noise about large language models (LLMs) and them possibly replacing programmers in the near future -- a perspective shared by many technology enthusiasts but met with a tad more skepticism from researchers and software engineers. OpenAI's release of ChatGPT has generated lots of discussion on the subject. Although complete automation might be far off, some folks have already found some use for Copilot, automating things like rote boilerplate code. However, that does not necessarily translate to correct, let alone efficient, code especially in complex applications and specific domains. An obvious caveat: different language models lead to different results, and the randomness in the inference process can create a lot of variability -- some friends have told me OpenAI playground does generate more reliable code.

Our paper Constrained Differential Dynamic Programming: A primal-dual augmented Lagrangian approach has been accepted at IROS 2022.

Check out the preprint here.

Consider once again an optimal control problem (OCP) in continuous time: $$ \begin{equation} \begin{alignedat}{2} &\min_{x, u}{}&& \int_0^T \ell(x(t), u(t)) dt + h(x(T)) \\ &\suchthat&& \dot{x} = f(x, u), \quad t \in [0, T) \end{alignedat} \end{equation} $$

Our paper Implicit Differential Dynamic Programming has been accepted at ICRA 2022! Check out the final version here.

In general nonlinear optimization, we seek to solve problems of the form $$ \begin{equation}\tag{NLP} \begin{aligned} \min_z{} &\ell(z) \ \suchthat &c(z) = 0 \end{aligned} \end{equation} $$ This kind of formulation is often encountered in machine learning, control problems and others where structured, constrained optimization crops up.

The Differential Dynamic Programming (DDP) is a nowadays quite popular optimal control algorithm, which was introduced by Jacobson all the way back in the 1970s. Recently, it has made a resurgence following work by Tassa and Todorov.

The goal of optimal transport problems, is to find optimal mappings between probability meaures: these mappings are also called transport plans, and can take the form of functional transforms (in Monge's original problem) or joint probability distributions (in the Kantorovitch relaxation).