Sagar Patil

I can’t stop taking pictures in AC Shadows!

Yes, I do not know how the use of latent actions helps when they actually deploy. Either they are completely different mode not used at all,or they have negative influence on the whole policy since these actions are not GT and are much coarser level, and obviously the frames can be unfaithful.

Dude, the cars below are overtaking my plane!!

Beautiful paper, and beautiful visaluzations!

A big thanks to the @naverlabseurope.bsky.social Spatial AI team for this, but in particular to @steevenj7.bsky.social, not only for the great work linking e2e training to a Kalman filter but also the D3.js magic making the real-time analytics work, and the interactive website!

One way of viewing bitter lesson is that we value convenience more than optimality.

Omg, thank you so much!

Please repost to get the word out! @nkgarg.bsky.social and I are excited to present a personalized feed for academics! It shows posts about papers from accounts you’re following bsky.app/profile/pape...

Generative models are blobs. Good, huge blobs!

Yes, I agree. I am thinking of a distinct future when we have all the resources in our solar system. Lol, too hypothetical!

Fine-tuning is a fine game when you know the test-distribution and I believe that active fine-tuning at test time will be on the cards. But none can truly generalize out-of-distribution. You either collect a hell lot of data or stumble upon an architecture that truly generalizes of OOD for your task

Example the bias of CNNs to exploit high-frequency components in image might be good for CV benchmark performance, if you feed the same representations to a robotic policy then it is clearly sub-optimal, shape is more important here ....

Well, is it efficient, yes! But is it the optimal, probably no. The problem with pre-trained modules is that their representations are biased by their objective to fit a benchmark. And this objective encapsulates years of community based benchmark overfitting that we don't even think of biases ..

I think pure end-to-end learning will be a deployment scenario. Most of the model based planning methods and priors will be used for data generation in simulation for efficient exploration in data collection. But it is hard to win over pure end-to-end learning system if you have the data once.

Kantha quilt involving multiple shades and hues of blue. The quilt is sitting on top of a bed.

Kantha quilt involving various shades and hues of blue along with red. The motif shows flowers among the patchwork.

Kantha quilt involving various shades and hues of blue, red, and white. The motif is made up of various triangles.

Kantha quilt involving various shades and hues of brown, blue, and red. The motif is made up of various squares.

love these kantha quilts and bedspreads at 11.11 / eleven eleven.

Yes, compared to previous years it has become much more accessible and more like typical supervised learning, which I believe is great. Also looking forward towards your lab opening and following your research. Daniel Brown talks about you!

Interesting,I think most people don't apply robotics as an application area because of slow feedback loop. I also don't think that architectures developed for other applications may not be the most optimal for robotics. I believe it depends on the assumption you are making on the data distribution.

New here? Interested in AI/ML? Check out these great starter packs!

AI: go.bsky.app/SipA7it
RL: go.bsky.app/3WPHcHg
Women in AI: go.bsky.app/LaGDpqg
NLP: go.bsky.app/SngwGeS
AI and news: go.bsky.app/5sFqVNS

You can also search all starter packs here: blueskydirectory.com/starter-pack...

Hmm, interesting. What do you think is the use case of RAG in robotics? Like Behavior Retrieval?(arxiv.org/abs/2304.08742)