MCMC Motion Planning for Boltzmann-Rational Trajectory Optimization

Jovana Kondic, Stewart Slocum, and Olivia Siegel

At a High-Level

Markov Chain Monte Carlo (MCMC) algorithms to sample diverse, nearly-optimal trajectories for motion planning
Experiments on a 2D navigation and 3D manipulation problems
Second-derivative information helpful for sampling in 3D manipulation problems
MCMC motion planners may help robots better model humans and improve human-robot collaboration

To model an agent, we can define a probabilistic model

p(\tau | g)

To model an agent, we can define a probabilistic model

p(\tau | g)

A popular model of the behavior of an agent with goals is that of Boltzmann rationality.

To model an agent, we can define a probabilistic model

p(\tau | g) \propto \textcolor{red}{e^{-\beta C(\tau, g)}}

A popular model of the behavior of an agent with goals is that of Boltzmann rationality.

It models a kind of approximately rational behavior.