Abhinav

@Scav everyday

Is anyone actually still using Obsidian?

@alpercanbe yeah that's why i decided to write this, would love your thoughts abhinavjha.xyz

every robotics paper has a flow matching section where they introduce flow matching with fancy little symbols

@_rajanagarwal whether the data used provides enough signal to differentiate between different plausible solutions to uniquely identify the desired one, is a question i don't think i've been asking before, really good read!

i don’t usually share when things go wrong but recently i was researching world models and it went sideways, but i learned a lot i wrote a new essay about how not to do research rajan.sh/multiplayer

@sighjith @MaxDiffusionRL rlly appreciate it, love the stuff you've done!

@MaxDiffusionRL @_AbhinavJ no but this is a great writeup!

i spent the last 4 days diving deep into flow matching and visualizing it inside vision-language-action models turning pure noise into coherent actions for robots to follow is beautiful here's the blog I wrote about it with visuals that made it click better for me:

@ak_cozmo the demos from @physical_int seem to show that they hold pretty well, do check out their blogs they're written so so well

@_AbhinavJ flow matching for VLA models is a game changer - converting noise to robot actions is such a clean way to think about policy learning. reminds me of why diffusion took off for image gen. how's the inference latency holding up for real-time robot control?

thanks! you're completely spot on, it's a mistake from my end. In the paper for π0.5 they denote t=1 as noise and t=0 as being inside of the target distribution. I personally think it's easier to interpret it going from t=0 (noise) to t=1 (target distribution) instead, which is why the equation in the first section was also x_t = (1-t)x0 + t*x1. But when the graph was plotted it was indeed from t=1 to t=0 since it's hooked up to the actual implementation of pi0.5 and I forgot to reverse it. It's been fixed now, great catch, and thank you for reading attentively!

@_AbhinavJ Great post! I’m a bit confused by the cosine similarity and L2 distance plots, they look reversed along the x axis, eg shouldn’t cosine similarity start at 0 then increase to 1? Or am I interpreting it incorrectly

@novasarc01 thank you, I love your blog posts!

@_AbhinavJ cool blog abhinav!

@devenshah2018 glad you liked it!

@_AbhinavJ love the interactivity!

@nicoup thank you

@_AbhinavJ great work!

@pham_blnh thank you, loved the stuff you guys did with the data collection pipeline on the humanoid

@_AbhinavJ love it

@_rajanagarwal thank you, really appreciate it

@_AbhinavJ really well written!

@JieWang_ZJUI @yacinelearning @nrehiew_ @KyleVedder @chris_j_paxton thanks! yeah I did, just wanted to get the first version of the blog out, will be adding that soon as well!

@_AbhinavJ @yacinelearning @nrehiew_ @KyleVedder @chris_j_paxton Congrats! Have you considered visualizing the 3D trajectory process with pi0 in the simulator? Could be a cool proof of concept.

@yacinelearning @nrehiew_ @KyleVedder @chris_j_paxton @JieWang_ZJUI it has to be these visuals that show how noise transforms into actions that seem to have some local structure

@_AbhinavJ @nrehiew_ @KyleVedder @chris_j_paxton @JieWang_ZJUI nice man congrats what’s the main highlight?

@nrehiew_, @KyleVedder, , @chris_j_paxton, @JieWang_ZJUI, @yacinelearning would love your thoughts on this

@alanxue_ as messed up as this is, this is so so sick

I built gambling for Clash Royale.

@nrehiew_ just wanted to say your blog on flow matching was amazing

Just noticed that my blog post on Long Context is cited here 🫡

@_advaitpatel has to be cli, it just works better, though it does sound robotic asf

when people say codex is better than claude code, are they referring to the CLI tool or to the async web version?

@elliotarledge It was such a big promise, completely forgot it was even made

The one thing we couldn't seem to get was 10 million context windows.