Michael Surazhsky

I work on AI that works in the physical world, to make it better.

[Michael Surazhsky @M_Surazhsky]

@jiriknesl This applies to more then you think, concepts across very different domains at first look

[Michael Surazhsky @M_Surazhsky]

@chris_j_paxton Supervised learning can only be as good as the data, unsupervised learning is limitless

[Kashu Yamazaki ✈️ CoRL @kashu_yamazaki]

I believe in simulation. It will solve everything eventually.

[Michael Surazhsky @M_Surazhsky]

Playing a lot lately with training loco-manipulation, here is a blogpost I wrote to compare the different data types used for training.
If you're in robot learning and deciding where your next hours of data should come from, I hope this map helps. https://t.co/h7hakD2wsX

[Michael Surazhsky @M_Surazhsky]

@QwQiao That didn't age too well...

[Michael Surazhsky @M_Surazhsky]

@amypretzel Engineering is about solving problems under constraints - which can be summarized as an RLVR task... research is more about decision making - where the solution is unknown or might not even exist

[Michael Surazhsky @M_Surazhsky]

A legged robot learned to walk, trot and bounce from minimizing only torque times joint velocity. No gait library, no reference motion, no reward for electricity ( the authors couldn’t measure it CoRL 2021, Unitree A1).

This is kinda like pricing your car’s fuel by engine RPM alone, nothing metering what idles under the hood.

Someone has now measured the real thing, on a different Unitree robot’s arm. Actual torque work was 10% of the electrical power draw. Copper loss was 42%. Friction, another 43%. The robot standing still, doing nothing, already pulled 135 W.

The proxy that produced a real gait never saw the terms that run the battery, as robots will start moving into deployment real energy spend would be harder to ignore.

Paper: Fu, Kumar, Malik & Pathak, CoRL 2021, arXiv:2111.01674
Paper: Deniz, Vega, Parsons & Auat Cheein, arXiv:2606.15915

[Michael Surazhsky @M_Surazhsky]

@sarahcat21 Founders who had hobbies before the startup or manage to keep them also during👀

[Michael Surazhsky @M_Surazhsky]

@martin_casado It's more like founders stop doing the hobbies because they focus on the startup, if they enjoy it it's not too bad

[Michael Surazhsky @M_Surazhsky]

@heyyalexwang So you view RL reasoning training more as imitation learning, and not real RL with exploration? why couldn't RLVR be used to check every reasoning step?

[Michael Surazhsky @M_Surazhsky]

As Anthropic and DeepMind rolling out their new models with a major progress in robotics people are saying that specialized model or LLM alternatives are doomed.

Scaling laws work - if you continue to give the transformer more rich and diverse data it will learn it's distribution, yes even in robotics.

This doesn't downsize the training and data Fable 5.2 and Gemini Robotics 2.5 used to train. And it doesn't mean that a slow demo of solving a task means it will translate to a reliable and fast execution that will survive deployment.

[Michael Surazhsky @M_Surazhsky]

@maximumbaelz @leothecurious We mostly don't understand the brain yeah, reliable in judgement, not necessarily performance. We might not know how we do it, but we can imagine outcomes before acting and choosing the best aligned action, this is not yet solved in statistical models

[Michael Surazhsky @M_Surazhsky]

@yunchanyq6 Yes, but not all tasks can afford recovery

[Michael Surazhsky @M_Surazhsky]

I suspect VLAs could be scaled the same way LLMs do, considering recent frontier ones even explictly trained on robotics data.
Figure’s Helix 2.0 allegedluy chained 61 actions over 4 minutes in a real kitchen, as ractive sequencing with no explicit lookahead.

Question is if the context managing we do with agent harnesses is enough for physical environments, and whether the representation of the current context is even correct for the physical world.

[Michael Surazhsky @M_Surazhsky]

@nelntz זאת דרך נחמדה להגיד שהם לא מספיקים מאמינים בך ואם בטעות תגדל אקספוננציאלית אז דבר איתם

[Michael Surazhsky @M_Surazhsky]

@rand_longevity biology will solve aging, we're not there yet.

[Michael Surazhsky @M_Surazhsky]

@maximumbaelz @leothecurious For knowledge work LLMs might be superior, but they are still statistical models, not perfect for decision making

[Michael Surazhsky @M_Surazhsky]

@leothecurious Well, not-brain-like architectures might lead us to solve brain-like architectures. We never had tools to compress and progress in distribution of human knowledge like this before

[Michael Surazhsky @M_Surazhsky]

@PraneetKedari As a lab it’s a generalist vs a specialist game, focus is your moat. But honestly end to end products will be the winner

[Michael Surazhsky @M_Surazhsky]

@DeryaTR_ Yeah, no.
Physical AGI is when you tell a humanoid to do a random task in an unseen env and it zero shots it.

[Michael Surazhsky @M_Surazhsky]

@cgtwts This isn’t really language anymore, tokens are very sparse in these model and the modalities are very diverse - it was clearly trained on robotics and 3d.

Anthropic bet on everything could be code really paid off, but representation isn’t the only problem in AI