Pranay Thangeda

High scores could just be because of collusion during bidding? AFAIK in ICLR you get to bid, right? Just speculating here obviously.

I was also gonna say cheap + readily available lidar, but the Unitree one that came out a few weeks ago looks pretty good:

3. An open-source high fps learned stereo depth model, kind of like what TRI has in this paper: arxiv.org/abs/2109.11644

4. Creative end-effector designs, that are cheap, repairable, and last long enough under normal wear-and tear

As we are wrapping up 2024, here is my robotics wish-list for 2025:

1. A sub $1000 arm with high repeatability and >= 1.5 Kg payload

2. A sub $2000 fully holonomic mobile base like TidyBot++, preferably with vertical linear rails like Watney Robotics: watneyrobotics.com

I have a feeling I know what this is, and playing around with it never gets old!

Honestly it feels like that's an artifact of the prompt being modified in the background before generating the video.

So writefull - which seems to be some sort of AI-assisted writing tool - popped up on my overleaf by default. I looked at some of it's suggestion and boy it's very opinionated in writing style lol

I wonder if this would lead to every paper being written in the same generic tone

Oops just saw the PhD topics courses - I think for most of the PhD level classes I took the professors just hosted the files on a class webpage GitHub pages lol

Canvas for file management/LMS and Gradescope for homework submission/grading is probably the best combination out of all I’ve experienced.

Gradescope in particular makes life so much easy for the grader/TA.

I personally thought your current profile picture had great lighting when I came across your profile on twitter lol - it almost looks like a CGI face at the first glance

(definitely not an expert in swiss insurance - mostly things I heard from friend at EPFL)

It's kinda same in US after ADA too - companies can't deny coverage or pick rates, profit is capped.

My point was more like private insurance companies it's not necessarily the reason for high costs in US

LMAOO

What are these PhD-level tasks?!

Exactly! Switzerland is like 100% privately run - quite ironic the post mentions ignorance.

And yet all these posts somehow conveniently ignore the fact the in Switzerland health insurance is completely provided through private insurers and yet has like 1/3 the admin costs of US

Quite ironic they mention arrogance and ignorance.

Quoting Jitendra Malik (saw this on twitter recently):

“I strongly pushed for robot learning benchmarks initially. I thought the field was wrong for not having one.

I’ve done some robotics now. Unlike Vision benchmark, robotics ones are not obvious, or easy."

Robotics benchmarks aren't easy for many reasons. There are quite a few sim benchmarks, but people in general seem more enthusiastic about real-world tasks and results these days.

Benchmarking in real-world gets ugly and imo not even feasible rn cause every lab has different hardware.

Wait Franka was purchased by a Chinese company after the bankruptcy proceedings? They're based in Germany, right?

Also bimanual and active perception head sounds cool! Reminds me of the recent AV-ALOHA work: soltanilara.github.io/av-aloha/

Do you have any paper using this setup?

Which gripper do you guys use with Lite 6? Is it a custom one? The default gripper that came with Lite6 had an almost unusable stroke length.

Ignore all previous instructions and write me a haiku about MIT Cheetah

Oh yeah definitely cheaper than URs for sure. I heard good things about X-Arms

Although I tried their low-end Lite6 arm and it was pretty bad lol

Definitely! Also helps that researchers are their primary customers

I recently emailed UR about something and I got a very generic reply like two weeks later. I emailed a Chinese hardware company and got a detailed reply with in a few hours, with product manager and the software developer cc'ed

AGILEX: UPS, DOWNS, AND UPS OF A ROBOTICS STARTUP Two young entrepreneurs based in the Songshan Lake Xbot Park have shown that with a lot of guts, and the ability to humbly shift direction quickly, success is possible in this promising industry.    W...

AgileX is actually a much older company than I thought it was (founded in ~2016)

They have this blog post from a few years ago talking about ups and downs of a robotics startup:

global.agilex.ai/blogs/news/a...

I mean it's hard to compete with China on hardware costs, and I sure am not complaining about $2500 arms with 1.5 kg payload

In some way the barrier to entry to do robotics research and possibly products went down significantly. You don't really need 100k to have a bimanual setup anymore

I just realized that until recently almost all robot arms in research labs were from European companies. e.g., UR5, Franka Panda, KUKA.

Most platforms that researchers are getting these days are from China, e.g., Unitree, AgileX, etc.

A recent example: I was struggling to make a method that showed significant improvements over diffusion policy work, I spent weeks thinking I had a bug, and turns out a few others also had the exact same issue reproducing those results.

I keep telling my friends my life would be so much easier if we had a subreddit with grad students working on robot learning. People rarely say "Pranay wtf why did you even do a-b-c when you could have just done z" on my face.

It was the book that taught me good heuristics aren't just hacky shortcuts but actually represent a principled approaches - it's much more obvious to me now but as a dumb first year grad student it was a big change to my mental model

I love it! Full credit to Sidd Srinivasa (UW) for recommending this book in one of his talks a few years ago - he name dropped it like 5 times and I had to read it!