This is mainly a copy/paste of notes that I prepped for a friend back in 2015, with updates now and then. Part 1: General stuff Takuhaibin (aka kuroneko) offers a relatively cheap service of sending your luggage ahead. You can send it to specific addresses, and have it arrive on specific days, specific times! Very handy for doing side tours. Ask your hotel or nearby 7-11 for help arraging it. There is often no soap in public restrooms. You may want to bring hand-antiseptic. Often people speak surprisingly little English. Keep that in mind when booking stuff (better to ask hotel lobby to book stuff for you rather than just showing up). use hyperdia to check train schedules. NOTE! With the JR pass you can't use the Nozomi/Mizuho/Hayabusa shinkansen, so uncheck that option! Hyperdia has been discontinued :( mini-souvenirs! Giving little gifts is a huge thing in Japan, and it's very convenient to have mini-souvenirs to give back (especially when people go waaay out of their way to help you, which happens more often than you might expect). My recommendations are these 25g ragusa: cheap, blocky (don't break) and small enough to carry around. Might melt in summer though... Untangling my Chopsticks is a nice, relatively short read if you want to get an idea of the culture beforehand (focused on kyoto and food). I would recommend getting a physical book and not kindle though, since at the end of many chapters are recipes. Go to at least one onsen! Onsen is a hot-spring bath (usually public but sex-segregated)Sentos are also super nice! (almost same as onsen, but the water isn't from a natural hot spring). Especially in the south, sentos are often still part of people's daily routine, and quite an experience.

Pick one afternoon a week, to spend primarily reading (not UROPs). I know I said you should be efficient with reading papers, but honestly, at the beginning it takes just reading a lot to get the experience of reading papers. Better to read a bit more than too little. Absolutely read first (and immediately) the Whitesides paper. The rest, you should actually prioritize papers that are relevant to your own research, that you find yourself. But when you’re not sure what to read, skim this list and pick something that sounds interesting: if not directly relevant, at least it’s a somewhat curated list of papers I think are interesting. Tips for reading papers Check first to see if the project has a video, which usually summarizes it nicely. Decide based on this if it is worth digging deeper.Make a first pass to get the high-level gist: read the abstract, the figures and captions, then skim the paper and look for the problem setting and contribution: these are typically at the end of an introduction and end of the paper, respectively. Sometimes these aren’t made explicit, that’s the authors’ fault. Decide based on this if it is worth digging deeper.Don’t be bothered if you don’t understand all the details, even after multiple passes; just make sure you understand what you need.Don’t assume the authors haven’t made mistakes just because the paper is published. Favorites Design of Materials and Mechanisms for Responsive Robots

Next [ ] Robot Model Identification and Learning: A Modern Perspective Paper list [ ] On the Design of LQR Kernels for Efficient Controller Learning [ ] A Tutorial on Energy-based methods [ ] Combining physics and deep learning to learn continuous-time dynamics models [ ] ENERGY-BASED MODELS FOR CONTINUAL LEARNING [ ] Can Direct Latent Model Learning Solve Linear Quadratic Gaussian Control? [ ] Geometric Robot Dynamic Identification: A Convex Programming Approach

Overall, we're listing here some things we know we would like to improve/learn, but even more helpful will be the things we don't know we could be more efficient/effective. Notes tensors of scalars (i.e. tensor because of the number of envs, but each quantity is a scalar) should be of shape [n], not [n, 1]. They shouldn't be unsqueezed. If the quantity could be a vector (e.g. commands, but maybe you currently only have 1 command), then it should be of shape [n, 1]. General How do sub-modules in git work? Checking-out/committing etc. still seems completely decoupled, are we doing something wrong, or do we just live with this?

Humans have a rich repertoire of motion skills, and are able to quickly recombine them in unexpected ways, for example to adapt to a new situation (parkour or soccer player) or even for aesthetic creativity (dancers or skateboarder). Robots, on the other hand, are still largely trained for specific tasks, and with a rather rudimentary metric on performance: largely effort-squared and some proxy on stability. A critical limitation, both in conventional approaches such as model-predictive control (MPC) approaches and reinforcement learning (RL), is a heavy reliance on reference trajectories. A key to recreating the similar richness in motion as humans, is the ability to reason across different scales and resolutions of both (state-) space and time, which we believe requires a (flexible) hierarchy of models and policies. We distinguish between horizon length: how far into the future you predict