ameyasalvi.bsky.social
PhD candidate in Robotics and Autonomous Vehicles. Interested in all things science, education, and engineering. ameyarsalvi.github.io
130 posts
1,502 followers
973 following
Regular Contributor
Active Commenter
comment in response to
post
I am unsure about “no value” part, but as generating more and more papers is getting easier (at least in quantity), I would believe a proportionally diminishing value they bring (a purely economical analogy)
comment in response to
post
Got a cool narrated video too
youtu.be/vrpJjbzV0mk
comment in response to
post
A few week backs I thought people's attention is diluted jumping in and out of several socials and thought soon enough that will settle down -- but maybe that's gonna persist a while.
comment in response to
post
All AI tools, vehicles is an example. Coding/ Writing assistants, all types of AI assistants, etc.
comment in response to
post
Fair, economic projection models have always been questionable I guess.
comment in response to
post
Yup! I turn it off because I really like writing (being poor to afford the full version is beyond the point 🥲) — but I see a future where all papers are nothing but templates filled with AI generated words
comment in response to
post
I agree, I am unsure about the scaling of the commercial aspects, at least in the passenger vehicle space, freights being a diff story.
I was more pressing on the viability of the AV tech. I do think tech is mature, but regulatory commissioning (rightfully) can put off OEMs.
comment in response to
post
I believe its to encourage industry contribution— there focus is more on codes/libraries/containers and much less on math and analytics. Probably, thats my guess…
comment in response to
post
Also we often use the NVIDIA products for several other things in our lab
medium.com/@dhruvm_6460...
comment in response to
post
I have used the Xavier occasionally for segmentation/ control. Its decent -- but have experience that the power consumption is a bit too much.
Although in prototypical projects algorithmic optimization for power consumption is rarely the focus.
comment in response to
post
Fantastic!! :)
comment in response to
post
Dr. Strogatz, do you have your available online?
Or is this available in any of the books you have authored?
comment in response to
post
As in they criticize things for lack of novelty? — cause they fail to understand that robust deployment if simple frameworks can be significantly challenging as well?
comment in response to
post
Lol, I agree.
Especially when academic funding flows in the direction of “hype”-cycles, everyone got to rebrand themselves as the pioneers of the next-big-thing.
comment in response to
post
A significantly similar argument in engineering is between model based control (which is still very arts-y with coming up with basis functions, modeling,etc) vs using ML (IL/RL) for policy learning.
Luckily the challenges with embodiment have kept the MBC relevant for now!
comment in response to
post
A significantly similar argument in engineering is between model based control (which is still very arts-y with coming up with basis functions, modeling,etc) vs using ML (IL/RL) for policy learning.
Luckily the challenges with embodiment have kept the MBC relevant for now!
comment in response to
post
More than actual articulation in words, the narrative is being reflected in actions — There is less and less reading of tech docs, stack exchange arguments and more and more of “give me a c++ code for this…”
And yes, pretty widespread. Almost every TA in school saw see this in coding assmts.
comment in response to
post
I bet it can’t run to the lobby after knowing about free food.
comment in response to
post
As a grad student, I see several grad students using GPT for coding, brainstorming, writing etc. Inarguably there has been a striking difference in the progress of students working with GPT vs without GPT.
If $$ becomes a factor in research progress, this is simply going to be an unfair advantage.
comment in response to
post
They slightly brushed on ‘modularity’ vs ‘end2end’ . They raised that modularity is more due to the fact that we are doing research in silos/ focused academic depts ( ex perception research/ control engg research, etc).
Modularity will seem to remain for the foreseeable future due to this.
comment in response to
post
Documetation. The dreaded documentation for all the failed parameters :(
comment in response to
post
(6/6) Closing Remarks: Arguably there would be a sweet spot where both the approaches co-exists (hopefully) with task/robot specific mainstreaming of one or the other approach.
comment in response to
post
(5/6) Against : Trust and safety guarantees are essential for safety critical applications and will become more prevalent as the the number of robots in physical world increase. Such certifications at present cannot be provided for majority of Gen AI methods.
comment in response to
post
(4/6) For : The trends indicate rapid miniaturization of compute capabilities with increase in performance and reduction in costs. These trends indicate the physical tech keeps on getting better and cheaper, thus enabling low energy compute.
comment in response to
post
(3/6) Against : Energy consumption and carbon footprint are one of the prominent concerns for Gen AI models. At present the current frameworks are unsustainable and are nowhere near the energy consumption standards of traditional model based approaches.
comment in response to
post
(2/6) For : There are several instances of the physical world that are extremely difficult to model (soft body interactions, contact mechanics, etc.). Gen AI would be a huge plus to tackle such situations to generate a wide variety of interactive scenes for learning from interactions.
comment in response to
post
Hahahah! The pleasures of life ✨
comment in response to
post
I was never a GPU expert. After writing a piece of Python-logic I ask GPT to optimize it for me. It does a really remarkable job.
For one my RL trainings the time went down from 60+hrs to less than 20hrs
comment in response to
post
Whatever is upvote equivalent here ⬆️
Pursuit of perfection can sometimes be your worst enemy…
comment in response to
post
Seems very ‘kimchi pancake’ …
comment in response to
post
Transfer learning for sure! At least in the context of DRL for robotics