We want to cover interesting papers with a focus on those recently published. Everyone is expected to read the papers on their own, so that the meeting can focus on discussion of the paper's merits, issues, and possible extensions.

While we'll have a nominal queue of papers to follow, we should be open to changing what we read and discuss the following week, based on our discussion of the current paper.

Each week one member of the reading group will be designated as the moderator. Your job should not be to present the entire paper to the group, but instead to encourage participation by summarizing important parts of the paper and having probing questions ready to excite discussion.

We would like to pick our papers from recent robotics and machine learning conferences such as RSS, ICRA, IROS, ICML, NeurIPS, and Humanoids. Journal articles from IJRR, TRO, Autonomous Robots, JMLR, ML, and PAMI are also a good place to look. We have a more complete list of publication venues if you are interested.

We meet in the ll4ma lab, at 1:00-2:00PM on Thursdays (Spring 2024).

We manage meetings through our slack channel: '#reading-group'

Starting Spring 2021 we have added the following guidelines:

• We will have 2 designated discussion leaders
• We have a google document for people to share reflections on the reading prior to the group
• We maintain salient findings from the meeting in a single google doc
• We summarize every paper into a single slide here to help us maintain a big picture
• Treat the meetings as continuations of a series, try and keep notes in one place to bring conversation from previous weeks into the current paper discussion
• Discuss experiments!
  • Not just the results, do the experiments make sense?
  • Were the experiments well executed / measuring the important things
  • What would the ideal experiment have been?
• Use the questions below to help guide you when you don't have other things to ask.

List of general questions to consider in reading / leading the group:

• What type of problem/task does this paper address?
• Could you apply this approach to your problem? If not, how might it be changed?
• What would happen if you replaced some of the novel / more sophisticated parts of the paper with well established methods for planning/learning/control/etc.?
• Does the paper present an interesting machine learning method?
• What's the role of different kinds of uncertainty? (e.g. model uncertainty vs state/observation uncertainty; aleatoric vs epistemic uncertainty)
• Which other tasks could this approach be applied to?
• Where does this paper fit in the bigger picture of manipulation learning?
• Which other papers are closely related to this one?
• Which learned information can be directly transferred to different scenarios?
• Would it be useful for our setup to implement this approach?
• Is this approach equally applicable to grippers and dexterous hands?
• What are the key weaknesses of the proposed approach?

This reading group is a practicum, meaning we try out what we learn. Our BitBucket sandbox repository is for everyone to share in a common library for implementing and trying out new algorithms and problems.

https://bitbucket.org/robot-learning/ll4ma-opt-sandbox

• Cutting planes for integer and mixed-integer programs
• Online optimization
• Nocedal and Wright Ch 7 - Large Scale Problems
• Gradients - Finite differencing; complex finite differencing; automatic differentiation - PyTorch, TensorFlow, Casadi
• Other topics on optimization
• Simulated Annealing is an Interior Point Method
• The Kinematic Hessian
• Optimization Fabrics