
Research Scientist @physical_int Previously Senior RS @ToyotaResearch, Robotics PhD @MIT_CSAIL. Towards generalist embodied intelligence 🤖 Opinions my own!
Joined April 2018
- Tweets145
- Following228
- Followers439
- Likes486
Pinned Tweet
I’m super excited to share that I’ve joined @physical_int as a Research Scientist!
Katherine Liu retweeted
Replying to @VilleKuosmanen
We actually got a VLM to perform better on VLM benchmarks after VLA training. arxiv.org/abs/2602.01067v1
But mostly it’s because we co-trained with VQA data while the backbone was just pretrained. But we also showed that we were on par with the VQA trained version.
Katherine Liu retweeted
hiring a GM for research ops! looking for someone who can manage managers, build great teams, solve hard problems, and create systems that absorb complexity.
must also be fun. we're going to spend a lot of time together 😇
jobs.ashbyhq.com/physicalint…
Katherine Liu retweeted
our bottom line is robotics and we are incredibly driven to deliver useful robots everywhere
apply here to push the frontier of robotics: pi.website/join-us
Takeaway: treat 3D as an image, and geometry inherits 2D pretraining.
Joint work with @s1wase , @yolo24312431 , Dieter Fox, @ZakharovSergeyN, and @robo_kat at UW and TRI.
🌐 geyan21.github.io/UniFlow
💻 github.com/geyan21/UniFlow
3D should intuitively be really useful for policy learning, but natively 3D representations are tricky to scale computationally. UniFlow leverages point maps, which have become popular in recent years because they allow you to represent point clouds as images.
What if 3D could just be an image?
Introducing UniFlow: we treat 3D pointmaps as images, so they can use the same pretrained 2D priors as RGB, with no 3D encoder or camera extrinsics needed. A multi-modal diffusion transformer with per-modality modulation then fuses the RGB and pointmap streams. We found this works better than pretrained 3D encoders and stays robust under domain shift. It outperforms the Large Behavior Model (LBM) in sim and π0.5 on a real robot.
Come to poster #71 at #CoRL2026 today, 5–6:30 PM in Lone Star (3rd floor), JW Marriott!
geyan21.github.io/UniFlow
UniFlow was also a precursor to Flexi-Pi, which you should also check out :) flex-pi.github.io
1
1
131
I’m sadly not at CoRL this year, but I highly encourage folks to find @GeYan_21 at the poster session to talk 3D and scalable policy learning.
1
76
What if 3D could just be an image?
Introducing UniFlow: we treat 3D pointmaps as images, so they can use the same pretrained 2D priors as RGB, with no 3D encoder or camera extrinsics needed. A multi-modal diffusion transformer with per-modality modulation then fuses the RGB and pointmap streams. We found this works better than pretrained 3D encoders and stays robust under domain shift. It outperforms the Large Behavior Model (LBM) in sim and π0.5 on a real robot.
Come to poster #71 at #CoRL2026 today, 5–6:30 PM in Lone Star (3rd floor), JW Marriott!
geyan21.github.io/UniFlow
After LBM 1 and VLA Foundry, I wanted to find a place to keep pushing on the science of generalist robot learning at larger scales. I’ve really been struck by both the scale and thoughtfulness (these two qualities can be difficult to maintain at the same time!) here at Pi.
1
4
406
Image augmentation is widely used to make learned vision algorithms more robust. MobileVISTA is the natural follow up to our previous work VISTA, exploring how to augment demos in the image space for greater robustness to changes in robot base positions.
Mobile manipulation policies can break from base pose errors of just a few cm, which are common after navigation. Can we get pose generalization without additional demos? Introducing MobileVISTA: collect demos at one base pose, get a policy that works from many 🧵👇
Website: sdwistreich.github.io/mobile…
Paper: arxiv.org/abs/2610.07511
Katherine Liu retweeted
Proud to debut Walden Robotics, and share we've raised $300M in funding co-led by Toyota and Deviation Capital, with participation from NVIDIA, Boeing Ventures, Samsung, and others.
We're building general-purpose robots that continuously learn while performing real work, alongside people. Today.
youtube.com/watch?v=fewvZrck…
I’ll be at ICRA this year! Excited to visit Vienna and catch up with everyone. I’ll be giving a talk at the VLA Pipelines Workshop on Friday afternoon. Looking forward to talking robots :)
3
334
Katherine Liu retweeted
Submit your CoRL workshop proposal! This year @RLioutikov and I wanted to make the workshop more "workshopy".
Main changes are:
- Half-day events only
- Limited speaker slots
- Challenge- and participation-driven
- A post-workshop artifact (white paper, report, paper, etc.) summarizing the discussions
📣 CoRL 2026 Call for Workshops & Tutorials is live!
This year: workshops should be workshops again, not mini-conferences. We want formats where the community actually works together.
Half-day events | Nov 9 | Austin, TX 🤠
Details: corl.org/contributions/call-…
#CORL2026
Katherine Liu retweeted
Releasing RecGen: a collaboration between @ToyotaResearch, @toyota_europe, and @UvA_Amsterdam tackling a core 3D vision challenge: reconstructing complete multi-object scenes (parts, poses, textures, even occluded geometry) from just 1 to a few RGB-D views.
Trained purely on synthetic data, RecGen achieves SOTA on real-world robotics and 6D pose benchmarks, handling occlusions, symmetry, and complex interactions.
A step toward scalable, high-fidelity digital twins for robotics, and better evaluation and training of generalist policies.
reconstruction-by-generation…
Katherine Liu retweeted
a good model trained on even a simple task with a tiny amount of data feels mesmerising, no matter how many times you see it.
Katherine Liu retweeted
Gave a talk on Robot Learning from Human Data at Stanford. It was great to be back!
Some opinionated points:
1. Human data collection capacity is outpacing the research.
2. We still don't have the "science" for scaling robot capability with human data.
3. We are far from being able to model naturalistic human behaviors.
youtube.com/watch?v=NUtaN10c…
[5/5] But wait, there’s more! Bring your own datasets, load from LeRobot, visualize everything, and explore all tutorials 🧩: github.com/TRI-ML/vla_foundr…
Project website: tri-ml.github.io/vla_foundry…
[4/5] Grand finale: evaluate your VLA Foundry robot policies in simulation on 40+ challenging manipulation tasks! 🤖: github.com/TRI-ML/vla_foundr…















