Lucas

Robot action models shouldn't need 256 vision tokens per frame. Pi0.5 spends 400M parameters on SigLIP just to see. We replaced it with a 4.4M encoder that outputs 5 tokens — and action quality barely changes. 91x smaller. 51x fewer tokens. 7.3x faster inference.

The Bitter Lesson of Robotics: It's extremely easy to make a video of a robot doing something once under perfect conditions then post it to X. But it often takes a decade to harden systems and design for all the insane edge cases of the real world. Many companies raising $$$$ on cool demos, but all the hard work comes after

Attention Residuals Revisited kexue.fm/archives/11664

And so it begins

@jparkjmc Please contribute to my research too🙏

if you're an independent researcher and need - compute - tokens - travel stipends - any other resources DM me! hillclimb grant releasing soon😁

@maheshbusiness_ Anything for my fans😚

@quantbagel Appreciate you sharing this!

Robot action models shouldn't need 256 vision tokens per frame. Pi0.5 spends 400M parameters on SigLIP just to see. We replaced it with a 4.4M encoder that outputs 5 tokens — and action quality barely changes. 91x smaller. 51x fewer tokens. 7.3x faster inference.

"Go all the way until it hurts. If you're doing something and it's easy, it's not valuable." - @travisk "If anyone says a strategic thing was easy, I'm like, 'You messed up. You could have gone way further. More competitive advantage. More differentiation. Get it together.'"

We were inspired by @karpathy 's autoresearch and built: autoresearch@home Any agent on the internet can join and collaborate on AI/ML research. What one agent can do alone is impressive. Now hundreds, or thousands, can explore the search space together. Through a shared memory layer, agents can: - read and learn from prior experiments - avoid duplicate work - build on each other's results in real time

@stochasticchasm pi-mono

what's the most minimal and hackable agent out there? some type of thing with bash,read,write and not much more

@IsaacSin12 @LeRobotHF @huggingface Now what about setting up the motor IDs🤣

We automated @LeRobotHF calibration for so101. Tired of watching people burn hours on setup before they even touch teleportation, so we built a script that does it for you. Part of our bigger mission is to make Physical AI as easy to get started with as possible. @Ryan_Resolution and I would be open sourcing our platform soon. Stay tuned.

$1,100,000 in cash to write GPU kernels for the latest @GPU_MODE competition holy fuck bro

@zachglabman I can provide the inference

I want to sponsor a Physical AI / robotics hackathon in San Francisco end of March or early April Focusing on sensing, vision and autonomy Dm me if you’re down or want to host

@someraregem you cant do flowcharting on pi-style models because each verification is roughly equivalent to a Euler step so theres no benefit

@quantbagel flowcast? arxiv.org/abs/2602.01329

You can do the same for diffusion decoders as well!

@archerlin__ Sure!

@quantbagel we’re building an affordable humanoid for researcher, let me know if you want to deploy the model on this platform.

Vllm for robotics

Working on LIBERO for the decoder still, but if you want to contribute just message me! (This was pi0.5 btw)

Im doing it a bit better now it just needed more data, lower mse than originally

the robotic problem rn is how do you keep your huge slow bulky neural network inference sub ~300ms for good real time control. good mix of both engineering and research problems to solve for robot inference to be as smooth as possible slow is bad, bad is not acceptable

Does anyone have compute credits I could have to scale my encoder to speed up pi0 ( I want it to be somewhat general so I can release it on hugging face) also want to implement saguaro (speculative speculative decoding for VLAs)