Greg Tarr

making robots useful @markovrobotics

[atharva @_gundawar]

World models are very precise on out-of-distribution tasks!

Here’s an example of non-compliant arms running at low Hz, picking up a fake $100 bill. The policy was never trained on this or any remotely similar objects

Zero-shot generalization is the only way to physical AGI

[Greg Tarr @Greg_Tarr]

Play a video game and get a job. Winners get to work with real robots!

[atharva @_gundawar]

We made this game to hire cracked teleoperators (upto $100k and equity)

Play here: https://t.co/mvbSPTEH6T

Here is our friend Charis's run:

[Greg Tarr @Greg_Tarr]

Who said world models can’t be fast?

[atharva @_gundawar]

Real time chunking world models: 1x, autonomous, 200ms per generation!!

[Greg Tarr @Greg_Tarr]

@NGC6514 @_gundawar It is indeed @DexmateAI - with @SharpaRobotics hands.
The base model is LTX and we train on in-house data.

[Greg Tarr @Greg_Tarr]

@latentbunnies Always practice safe RTC

[Greg Tarr @Greg_Tarr]

My wife and I munching in the background

[atharva @_gundawar]

Egocentric view of RTC for world models

[atharva @_gundawar]

world model powered toaster!

[Greg Tarr @Greg_Tarr]

toast tastes extra good when made by your robot

[Greg Tarr @Greg_Tarr]

ltx ftw

[Markov Robotics @markovrobotics]

LTX-2.5 is changing how we train robots. Years of hand-built training data, replaced by a world model that already understands physical space. Trained on LTX. @ltx_io

[Greg Tarr @Greg_Tarr]

@_gundawar i think we’re going to make it (toast)

[Greg Tarr @Greg_Tarr]

@21nemus Thanks! I’m excited to see what we can do with 10-100k hours

[Greg Tarr @Greg_Tarr]

and we don't have any glasses in the ~4hrs of training data

[Chris Paxton @chris_j_paxton]

Picking up a fragile glass with robot hands

[Markov Robotics @markovrobotics]

0-shot pick and place on unseen objects with less than an hour of teleoperation data

[Greg Tarr @Greg_Tarr]

@AlanHollandCork @markovrobotics Thanks Alan!

[Greg Tarr @Greg_Tarr]

Excited to start sharing what we've been working on @markovrobotics

[Markov Robotics @markovrobotics]

Toward general dexterity

We are training robot policies that learn a rich and coherent physical representation of the world by conditioning on multimodal observations

Here’s a small glimpse of what we’ve been building:
Task: Pick the ramen cup and place it in the box

Given the current world state, the model generates a multimodal trajectory; here we show the decoded video and the corresponding actions executed on the humanoid

[Greg Tarr @Greg_Tarr]

@_gundawar jumpscare!

[Greg Tarr @Greg_Tarr]

@BenAybar @agi_inc The approach certainly does. We have the best computer use, mobile use, and web agents in the world.

[Greg Tarr @Greg_Tarr]

we @agi_inc just achieved 76.3% on the OSWorld benchmark taking the #1 spot from ByteDance (53.1%)

[Greg Tarr @Greg_Tarr]

@BenAybar @agi_inc I'm not sure that's public information right now. I can tell you that it doesn't matter enormously which base model one chooses as long as you have really good data and RL.

[Greg Tarr @Greg_Tarr]

@BenAybar @agi_inc Yes, it's a general model. It can perform coding tasks, but it's not optimized for programming - tools like Codex or Claude are much faster and better suited for that.

However, it can still assist in areas like automated UI testing when used alongside those specialised agents!

[Greg Tarr @Greg_Tarr]

Link: https://t.co/ZXVFxQLRVE