Starting a thread of recent talks on robotics, robot manipulation, and robot learning!

This is not an exhaustive list, just the ones I've watched and found insightful

Will keep the list updated as we go! Also looking for recommendations of other talks missing here

Starting off w/ a few fantastic talks from @RoboticsSciSys RSS2020. Here's the Test of Time award talk by @fdellaert and @michaelkaess on the past, present, and future of factor graphs and SAM

Keynote by Josh Tenenbaum on the importance of intuitive physics in robot perception and planning

Learned a lot from this talk by Byron Boots, who used online learning as a unifying perspective of MPC, RL, and IL

Talk by @lucacarlone1 on perception algorithms that are certifiable, real-time, and give semantic understanding

Talk by @leto__jean on why robot manipulation algorithms should continuously replan, exploit the environment, and carefully choose action spaces

Moving on from RSS, here's a talk last year by Russ Tedrake at @the_IAS on how keypoint-based manipulation enable generalization within object categories, and how to test manipulation policies for real-world deployment

Not directly related to robotics but this lecture by He Sun at Caltech's CS159 gives a great overview of different ways ML is used to solve "inverse problems" - inferring a data generating process from its observations. seems useful for sim2real!

Here are couple talks at @ipam_ucla's Intersection between Control, Learning and Optimization this year.

Talk by Richard Murray on ML in real-world, safety critical systems

Talk by Francesco Borrelli on how learning can be combined with MPC

Talk by @beenwrekt on quantifying perception errors for controls and giving robustness/generalization bounds

@RoboticsSeminar has been holding weekly robotics seminars with panel discussions - highly recommend checking them out! Here's a talk by Leslie Kaelbling on what's needed to build more intelligent and autonomous robot agents

And talk by @siddhss5 on a path-based perspective on search and how that informs edge-optimal search algorithms and Bayesian anytime motion planning