Greg Tarr (@gregtarr) header
Greg Tarr (@gregtarr) profile picture

cofounder @markovrobotics

San Francisco, CA
Joined September 2019
Photo by Greg Tarr (@gregtarr)Photo by Greg Tarr (@gregtarr)Photo by Greg Tarr (@gregtarr)Photo by Greg Tarr (@gregtarr)Photo by Greg Tarr (@gregtarr)Photo by Greg Tarr (@gregtarr)Photo by Greg Tarr (@gregtarr)Photo by Greg Tarr (@gregtarr)Photo by Greg Tarr (@gregtarr)Photo by Greg Tarr (@gregtarr)
i’m so proud of this
…so world models have x-ray vision?
2
1
234
to be clear, these are sub-cm 3d hand annotations using just a monocular view under full occlusion
86
I think this is a great take on the future of robotics
When I started my PhD, quantum computing gatherings were small. By the time I shifted to AI, they filled conference halls. CoRL this year feels the same. Here are some thoughts on what comes next for robot learning.
5
17
5
255
16,970
Greg Tarr retweeted
we need moar robot demos
For those of you DMing me, yes, we’re throwing another one at the house tomorrow, bringing together SF people in robotics, AI, and all things frontier tech! See you all soon :) luma.com/yeu7dx2f
2
1
8
551
at gemini robotics where an Apollo 2 robot is asked to pick up a donut from a box and put it on a tray. it grabs the whole box, tosses the donut, and misses the tray If this doesn’t show we’re early, not sure what does. #CoRL2026
2
4
21
1,169
Greg Tarr retweeted
.@markovrobotics is teaching robots to tackle complex manipulation tasks with fewer than ten hours of robot data. Their bet: video already contains most of the physical reality a robot needs to act. They use Modal for video processing, world model training, and remote inference.
5
7
2
48
5,359
Greg Tarr retweeted
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
6
9
3
111
108,399
Play a video game and get a job. Winners get to work with real robots!
We made this game to hire cracked teleoperators (upto $100k and equity) Play here: telebench.vercel.app/ Here is our friend Charis's run:
1
14
1,011
Who said world models can’t be fast?
Real time chunking world models: 1x, autonomous, 200ms per generation!!
3
573
My wife and I munching in the background
Egocentric view of RTC for world models
1
1
12
19,593
Greg Tarr retweeted
world model powered toaster!
super excited to partner with @ltx_io for a Day 0 launch with them!
8
11
1
86
18,831
ltx ftw
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
2
401
and we don't have any glasses in the ~4hrs of training data
Picking up a fragile glass with robot hands
1
7
1,080
0-shot pick and place on unseen objects with less than an hour of teleoperation data
1
3
1
20
3,624
Excited to start sharing what we've been working on @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
5
12
865
we @agi_inc just achieved 76.3% on the OSWorld benchmark taking the #1 spot from ByteDance (53.1%)
2
1
12
901
we're hiring as well, come join us!
AGI, Inc. is now the global leader on the AndroidWorld benchmark, with state-of-the-art verified performance of 97.4% This is a huge milestone for Android use, and just a sneak preview of what's coming - bringing trustworthy, reliable agents to every screen 🚀
5
683
As promised here's a TPA implementation in PyTorch w/ KV cache: github.com/Greg-Tarr/tpa-pyt… Didn't reference the authors' code so it might deviate a bit. Next step is to add a kernel to compute attn scores without materializing a ⊗ b in memory
6
516