Judging imperfect reasoning is a different skill from producing correct reasoning, and real student work is exactly where that gap shows up.
AI can solve a Millennium Prize Problem. It can’t grade a college physics exam.
On real handwritten answers, the best frontier model matches the instructor only 57% of the time.
After grading 6M student answers from 200+ universities, we’re releasing GradeBench (1/8)🧵
1
4
531
Fun side project cooking up: rich hand interaction data for steerability
Inspired by @maxtitovvv and @ElijahGalahad's blog, I've been digging into interactive steerability with hands.
The same hand motion can mean a light touch or a firm press. Would rich hand-object data with contact help robots learn that difference?
Sookwan Han @ CoRL2026 retweeted
Very early work, but seeing these things become possible, I feel like a pretty big paradigm shift in robotics could come soon.
LLM pretraining worked largely because the human data was already there, people just naturally writing about all kinds of things on the internet, not from collecting data for each task. Robot data today is kind of the opposite, people teleoperating robots one task at a time.
Collecting data for this project, I didn’t do anything special; we just moved around naturally in our house for a few hours, and suddenly the robot could just do the same interaction tasks we were doing.
My bet is that the data scaling for robots that live with people won’t come from deliberate collection, it’ll come from people just living their lives as in this work.
We see a lot of cool robot demos, but where are the people? Didn't we want robots that live and work with us?
The missing data: how people naturally interact.
We present Wearable-Free Human Real2Sim2Real, and demonstrate that the interactive abilities transfer well from human, to simulation, to real G1.
Reported in CoRL 2026 Late-Breaking Results 🙌
affordance.homes/lbr
Meet you in Texas 🤠
We see a lot of cool robot demos, but where are the people? Didn't we want robots that live and work with us?
The missing data: how people naturally interact.
We present Wearable-Free Human Real2Sim2Real, and demonstrate that the interactive abilities transfer well from human, to simulation, to real G1.
Reported in CoRL 2026 Late-Breaking Results 🙌
affordance.homes/lbr
Meet you in Texas 🤠
Sookwan Han @ CoRL2026 retweeted
👋 KRAFTON's Raon is entering the open-weight scene!
Meet Raon-1 mini, our first model, trained from scratch.
On par with leading models in its size class, while running faster and using less memory.
Available under the Apache 2.0 license.
huggingface.co/KRAFTON/Raon-…
Sookwan Han @ CoRL2026 retweeted
Today, we introduce Priming Models, a new paradigm for training one-step generative models from scratch.
No diffusion teacher. No flow map. No CFG. No JVP.
We believe this work will challenge how we think about building one-step models.
Joint work with @sitanch, @du_yilun, @parkwbin03, Bohyung Han.
🔥 1.19 FID one-step on ImageNet 256×256 while using substantially less compute, e.g., 12X less than Drifting-B, 6X less than Drifting-L
🎨 Scales to text-to-image generation (GenEval 0.82)
🤖 World modeling on RT-1 / Bridge-V2
Priming Model is built on one core message: [Reconstruction pretraining --> Distributional post-training].
Sookwan Han @ CoRL2026 retweeted
cool paper! beyond the nice results, hints that there might be better ways to train continuous-space generative models
Today, we introduce Priming Models, a new paradigm for training one-step generative models from scratch.
No diffusion teacher. No flow map. No CFG. No JVP.
We believe this work will challenge how we think about building one-step models.
Joint work with @sitanch, @du_yilun, @parkwbin03, Bohyung Han.
🔥 1.19 FID one-step on ImageNet 256×256 while using substantially less compute, e.g., 12X less than Drifting-B, 6X less than Drifting-L
🎨 Scales to text-to-image generation (GenEval 0.82)
🤖 World modeling on RT-1 / Bridge-V2
Priming Model is built on one core message: [Reconstruction pretraining --> Distributional post-training].
Simulation is great, except for human motion. It's hard to generate, and it rarely looks natural.
So, best of both worlds: simulation + real human trajectories
Sookwan Han @ CoRL2026 retweeted
1/8 Introducing Beckmann World Models (BWM): adapting Beckmann transport into fast, one-step action-conditioned world models.
Predicting the next video state from recent observations, proposed actions in one denoising step function call.
Our final RGB model predicts four PushT frames in 24.7 ms on one A800 GPU.
my take on deployment-first approach for manufacturing: makes sense, but if we're not creating the equipments (for welding, molding, etc) itself, do we really need that much of 'physical AI'?
2
630
Sookwan Han @ CoRL2026 retweeted
Introducing PAMI 🧩⚓: Part-Anchored Motion for Text-to-HOI Generation!
How should an object move with a human?
⚓ Anchor it! PAMI lets body parts guide the object with anchors.
🔗 coral79.github.io/pami/
📄 arxiv.org/abs/2609.38466
💻 github.com/Coral79/PAMI-Code
















