@GordonWetzstein

Professor at Stanford University & Co-founder at Rhoda AI

Joined January 2010
  • Tweets281
  • Following53
  • Followers5.6K
A secure grasp needs continuous feedback on contact forces. At Rhoda, we’re giving robots tactile sensing and teaching our models to predict what they’ll feel next—a step toward more efficient, reliable manipulation, even when vision is available.
Rhoda AI
@RhodaAIOct 14
What can a robot do without vision? We gave our robot a sense of touch and showed it can recognize and retrieve objects with its cameras blindfolded. Introducing Rhoda’s new tactile sensor—it feels pressure, texture, vibration, slip, and temperature, just like your fingertips.
1
3
1
39
1,967
🕹️ INSERT COIN. Spend your tokens wisely. Our Level-of-Token Diffusion gives image & video generation finer tokens where detail matters, coarser tokens elsewhere. Fewer tokens, faster generation. Enjoy the ’90s arcade vibes 👾 georgenakayama.github.io/lot…
4
5
54
3,209
Real-world robotics is where things get interesting: contacts are messy, objects move unexpectedly, grasps fail, and two hands often need to coordinate in very different ways. This is why I’m excited about the direction we’re taking at Rhoda: moving toward general robot models that don’t just execute learned motions, but perceive the consequences of their actions, adapt, and keep making progress toward the task.
Rhoda AI
@RhodaAIOct 6
Continuing our RhodaAtWork series with another look at Direct Video-Action (DVA) in action. This time, a single model handles multiple tasks across two home appliances: refrigerator EPS foam installation and washing machine drain hose routing.
27
3,719
Diffusion models spend the same compute on a blank wall as on a face. But you often know in advance where the detail will be. Introducing Level-of-Token (LoT) Diffusion: we turn that knowledge into a multiresolution token layout, with fine tokens where detail is needed and coarse tokens elsewhere. 1/9🧵
21
92
13
812
57,326
more replies
Generation adapts to the layout: fine tokens go where the prompt needs detail, and the speedup is set by the layout's token budget. Here LoT (right) uses 6,632 tokens instead of 14,336 and generates 2.5× faster than full resolution (left), and up to 4.6× when detail is more concentrated. 8/9
2
1
12
2,295
Work led by @GeorgeNaka40190, together with @BrianCChao, @jan_on_x, @HanshengCh, @fedassa, Leonidas Guibas, @YarivLior. 📄 Paper: arxiv.org/abs/2610.05816 🌐 Project page: georgenakayama.github.io/lot…

Level-of-Token Diffusion

Image and video diffusion models allocate equal computation to every region, even when the intended scene calls for varying levels of detail. The spatial distribution of detail can often be...

arxiv.org
4
6
53
2,617
The recovery example is especially compelling. A robot that can correct an imperfect action and continue makes a much wider range of workflows feasible. Recovery deserves to be a first-class capability in how we train and evaluate these systems.
Rhoda AI
@RhodaAIOct 1
More industries. More real customer work. Introducing RhodaAtWork, a series on the customer workflows our robots are learning to handle across manufacturing and logistics. First up: unpacking, scanning and stacking component reels for electronics manufacturing.
18
2,115
The most exciting result: a video-prediction metric measured before any robot post-training correctly ranks all seven models by real-world robot performance. This hints that progress toward better robot foundation models may be measurable—and scalable—in the world of video pre-training.
Rhoda AI
@RhodaAISep 15
At Rhoda, we care deeply about the science of pre-training for robotics. In one of the most rigorous studies of its kind, over thousands of trials and hundreds of hours of robot evaluations, we show how scaling web-video pre-training leads to better real-world robot performance.
52
5,853
This is a big step towards bringing robots outside the lab and into real-world environments where they perform useful tasks. The core insight of this blog post is that pretraining at scale, even with general web video, helps downstream robotics manipulation tasks a lot.
Tongzhou Mu 🤖🦾🦿
@tongzhou_muSep 10
Does scaling pre-training on general web video improve a complex manipulation task in real deployment? We scale model size and pre-training compute, and test on one industrial task. Yes. The better a pre-trained model predicts web video, the better its post-trained policy. 🧵
Seven pre-trained video-model checkpoints on one chart. Task performance on a real robot (at-speed completion rate: every requirement met within 100 s, 0 to 100%) rises with pre-training quality (DINO FD on held-out web video, log scale, axis reversed so better is to the right). XS, S, M, L are model sizes. The four M points are one model at 0.08×, 0.18×, 0.37× and 1× of its pre-training compute.

ALT Seven pre-trained video-model checkpoints on one chart. Task performance on a real robot (at-speed completion rate: every requirement met within 100 s, 0 to 100%) rises with pre-training quality (DINO FD on held-out web video, log scale, axis reversed so better is to the right). XS, S, M, L are model sizes. The four M points are one model at 0.08×, 0.18×, 0.37× and 1× of its pre-training compute.

1
2
27
2,846
Exciting results from Atlas, World Labs’ video-world model! Interested in its space-time simulation feature? Our CVPR 2026 paper BulletTime presents the technical details of decoupled time and camera control in diffusion transformers. 1/3 arxiv.org/abs/2512.05076

BulletTime: Decoupled Control of Time and Camera Pose for Video Generation

Emerging video diffusion models achieve high visual fidelity but fundamentally couple scene dynamics with camera motion, limiting their ability to provide precise spatial and temporal control. We...

arxiv.org
World Labs
@theworldlabsSep 1
Introducing Atlas: The world's first multimodal world model that generates image and video frames with pixel-perfect camera control and reconstructs them in 3D. Model the world, move the camera, and simulate space & time.
2
6
1
58
8,579
A key insight from BulletTime: frame index and world time should be separate. Giving the model explicit time and camera controls lets you freeze the action while moving the camera. Simply retiming the input before changing viewpoints can lose information and consistency. 2/3
2
5
687
A surprising finding: changing the positional encoding can outperform adding learned conditioning. In our time-control ablation, Time-RoPE alone beat standard RoPE + AdaLN on PSNR, adding no learnable parameters. Combining Time-RoPE with AdaLN worked best. 3/3
2
488
It's great to see my lab's reseach on CameraCtrl and CameraCtrl II in action for a scaled up video generation model. Great work - congrats!
World Labs
@theworldlabsSep 1
Introducing Atlas: The world's first multimodal world model that generates image and video frames with pixel-perfect camera control and reconstructs them in 3D. Model the world, move the camera, and simulate space & time.
1
93
9,603
Gordon Wetzstein retweeted
Kevin Zakka@kevin_zakkaAug 27
Powered by mjlab 😎
Pollen Robotics@pollenroboticsAug 27
Train in sim. Run it for real. Teach it new tricks. 🦆 This is the sim2real loop that powers Microduck. RL stack open source here : github.com/pollen-robotics/m… Buy it here : pollen-robotics.com/microduc… Git : github.com/pollen-robotics/m… Join our community : discord.com/invite/pollen-co…
3
13
171
15,932
Load more