Joe Resident

Wilbur Gardening Robot log - June 27, 2025 This week I surveyed the research on learning-based methods for robot mapping, localization, and navigation. I went this direction after spending a couple weeks surveying traditional SLAM (Simultaneous Localization and Mapping) approaches. There's a lot there. And then I realized, 'this is the old world. this was all made before LLMs, VLMs, and VLAs existed. I should focus on the new world'. So I switched. I think learned SLAM is the answer for Wilbur. A VLA is definitely the answer for actuation. After realizing that the evolution of traditional SLAM to learned SLAM would bring new generality, accuracy, and flexibility, I took some time to speculate about what the steady state of this research should be after all the fundamental advancements have been made. The 'Platonic Ideal' so to speak of learned SLAM. With my newfound clarity, I dove back into the research and was delighted to find that many of those ideas have been explored! So Thursday night and Friday were diving into the 'new world' of learned SLAM research, hoping to wrap that survey up by Tuesday, at which point I should have clear options in mind about what the weeding MVP robot should use. Then I'll probably narrow down possibilities and make a decision by the end of the week.

Reinforcement Learning will make super cheap hardware able to do very sophisticated things; the history of robotics has up to now been very focused on hardware and control, with actual outcomes and value delivered being afterthoughts appended to a primarily academic pursuit. But RL will make it increasingly about hardware and outcomes. Control will just be superhuman, without much engineering effort. We're gonna see so many cool robots in the next 10 years. https://www.youtube.com/watch?v=EZbM594T3c4

Exciting, I've been waiting for more attempts to use temporal dynamics in neural networks https://pub.sakana.ai/ctm/

Trump saying the Declaration of Independence was a declaration of "Unity, Love, and Respect" 🤣 Speaking of the document that basically says "That's IT, we've been oppressed too long, fuck you Britain, it's OVER, we're our own thing now, you're out" 🤣 "Unity, Love, and Respect" indeed, I'm dying, he says the funniest things sometimes https://www.youtube.com/watch?v=GBwCUPttprw&t=199

Thinking about using my real name for social media, including this account. Because I'm starting a public project, with my real face, and videos, and hardware, and don't want the mental overhead of hiding behind a pseudonym and vetting every post for identifying info. (May be related to my trace amounts of autism, lying or anything like it is extremely taxing. Not that I think pseudonyms are wrong, just that they require juggling multiple identities, similar to managing multiple realities when one has decided to lie about something). But, I'm also a privacy advocate and use a de-googled phone, encrypted email, VPN, etc. So it's against my knee-jerk tendency to maximize privacy. Good idea or not? #asknostr

Interesting paper I hadn't seen, the 'Densing Laws' of LLMs: they are getting twice as capable for the same size model every 3.3 months. https://arxiv.org/html/2412.04315v2 Qwen 3 released today may be an emphatic continuation of the trend. Need to play with the models more to verify, but the benchmark numbers are... Staggering. Like 4 billion handily beating a 72 billion model from less than a year ago https://qwenlm.github.io/blog/qwen3/