Srinath Mahankali @ CoRL

RL + robot learning phd @ berkeley

[Will Knight @willknight]

I wrote a feature for @WIRED about the fascinating rebels who believe the guardrails imposed on AI models are a dangerous form of censorship and control. By "abliterating" open models to remove refusals, these renegades want to make AI truly free--despite the dangers this brings. https://t.co/bsMBVo4s3B

[Richard Sutton @RichardSSutton]

Brute force computation is not enough for intelligence. Clever new methods are needed. The science of AI is far from finished. IMHO

[Yuval Tassa @yuvaltassa]

MuJoCo 4 is released!

https://t.co/j5X48JOQB8

[Srinath Mahankali @ CoRL @srinathm1359]

@NicoBohlinger this will go the way of nanogpt :D

[Srinath Mahankali @ CoRL @srinathm1359]

@ravisparikh this is so good!!!

[Srinath Mahankali @ CoRL @srinathm1359]

@yacineMTB pretty much

[Jagdeep Bhatia @ CORL 26 @JagdeepBhatia8]

Presenting at BARS tomorrow! Come chat at the poster @ 10:45 am.

[Srinath Mahankali @ CoRL @srinathm1359]

@LexnLin can you simulate lighting a match?

[Srinath Mahankali @ CoRL @srinathm1359]

@Bw_Li1024 does it see videos or just states + success or fail?

[Pulkit Agrawal @pulkitology]

@EkaRobotics has raised a $150M Series A.

Generality v/s Performance is the fundamental tension at the frontier of robotics. Generality without performance is just a prototype. Performance without generality is just a single-purpose tool.

Customers shouldn’t have to choose. Eka breaks the tradeoff. Powered by our Vision-Force-Action (VFA) model, we deliver the performance and versatility customers need to put robots to work every day.

Eka is built on three convictions:
⚡️Force is the language of manipulation.
🚀 Reinforcement learning is how dexterous manipulation scales.
🎂 System 1—fast, reliable physical execution—is the bottleneck to solve.

With this funding, we will scale our model and accelerate customer deployments.

Grateful to be building alongside a dream team and investors who share our ambition.

[Younghyo Park ✈️CoRL @younghyo_park]

Can a robot assemble tight-clearance parts using only its proprioception? No camera, no tactile sensor, no F/T sensor, no joint torque sensing, no priors over the object geometries or poses: just joint position readings.

It actually can! Blind policies trained in simulation insert pegs, mesh gears, and thread nuts, and the peg policy works zero-shot in the real world. Here's what we found 🧵

Joint work with @BronarsToni, my wonderful co-first author, who's on the job market right now! Advised by @kushalk_ and @pulkitology.

[Srinath Mahankali @ CoRL @srinathm1359]

@__0xhorror__ RL over the course of many human generations is really broad, generalization isn't a fact about RL but rather the curriculum distribution

[Srinath Mahankali @ CoRL @srinathm1359]

bitter lesson hitting robotics

[Octi Zhang @octi_zhang]

(1/6) From @GoogleDeepmind’s AlphaGo to AlphaZero, we saw machines find moves that reached beyond what we’d imagined for them. No one taught them those strategies. They emerged from reinforcement learning: playing, failing, and trying again under a single binary task reward for success or failure. Inspired by them, we couldn’t help but wonder: what might a robot discover if we gave it that same freedom to learn?

That question led us to AlphaNIST. Starting tabula rasa, with that same binary reward plus only a simple control cost, a single policy solves all 20 @NIST assembly tasks at a 98.2% success rate in simulation. More surprising still, AlphaNIST can complete 10 different tasks in a row, receiving that reward only once, after all 10 are done.

Just as AlphaGo’s strategies surprised its creators, AlphaNIST discovered dexterous strategies we couldn’t have come up with ourselves.

Paper: https://t.co/HZBYgG7SWK
Website: https://t.co/VyUKiYFI2y

[Srinath Mahankali @ CoRL @srinathm1359]

@waterloo_intern it probably depends on the number of such tasks + exact kind of backtracking computation being done, even though both require backtracking the concrete trajectory checking algorithm would be different so you need enough tasks to interpolate between backtracking algorithms

[Octi Zhang @octi_zhang]

(1/5) Everyone says dexterous manipulation needs real-world data. These policies were trained with none. Our robots mesh gears, thread nuts, and insert pegs zero-shot from RGB, and the same recipe works for locomotion across terrains much harder than before.

The missing piece wasn't a better RL algorithm, but a small tweak to what the robot practices.

Today we're releasing Success-Guided Sampling (SGS), a simple change to the task sampler that unlocks new levels of dexterity and agility with large-scale, end-to-end sim2real RL.

No demonstrations, no per-task reward tuning. Just PPO at scale.

Check out the policies yourself at https://t.co/dnwUr0tabh! All videos at 1x speed! 🧵 🔈

[Anjney Midha @AnjneyMidha]

hundreds of MI355x and B300 nodes

now live on https://t.co/t0uE0bekz3

subsidized for all .edu and .govs

it’s time to scale the frontier ecosystem

for the public benefit

[Seohong Park @seohong_park]

Doing proper science in robotics is hard, because human demonstrations are expensive and less controllable.

OCBench is a new benchmark providing "human-like" scripted policies. This enables scientific studies of BC and RL with virtually unlimited data!

[Srinath Mahankali @ CoRL @srinathm1359]

@Stone_Tao @Luckyballa the urdf is pretty good visually though, at least

[Srinath Mahankali @ CoRL @srinathm1359]

@zzigakovacic Amazing work! From the examples, it seems like Astra favors scripted motions while Opus favors more general simulation based solutions. What happens under perturbed conditions - how do agent solutions generalize?