Interaction with the real world is the major bottleneck in robot learning. So what would robot RL look like if we didn’t need to limit compute per interaction? Our latest work, Off-Policy Generative Policy Optimization (OGPO, accepted to ICML26) embarks on answering this question (spoiler alert: when done correctly, it helps massively!).
🧵(1/N)
Launching project flow. Our new generation robot hand.
Direct drive, liquid cooling.
From now, the heating problem is solved for robot hands.
more soon… @origami_robo
1/12 🚨 A long-standing goal in robotics is fully autonomous adaptation. Yet sparse rewards and weak zero-shot exploration often leave robots with little signal to learn from.
Our new work asks: What does it take to make autonomous improvement work with minimal human guidance?
Very fun collaboration on understanding how semantic generalizability is observed in multi-task VLA-RL for manipulation by scaling Q-function training.
RL post-training is often viewed as a sharpening mechanism that overfits to training task distributions and fails to generalize in OOD settings. But what if that's just a scale issue?
Atleast for semantic variations in simple manipulation tasks this seems to be case. Multi-task training improves performance on unseen scenes and objects with no policy updates (Best of N sampling via Q-values) and even more with policy extraction updates (OGPO, AWR)!
Can RL post-training of generalist robot policies enable improvement not just on the tasks we train on, but also generalization beyond them, improving on novel tasks not seen in training?
We show that, with the proper RL recipe, we can enable such "generalist improvement"! 🧵
introducing mjorbit: an open-source simulator for space robotics that runs thousands of worlds in parallel on a GPU.
here, a single RL policy trained in mjorbit flies in from kilometers away in LEO, grabs a satellite with both arms, and points it at earth 🧵
Back in ~Dec 2025, I really wanted to work on mobile manipulation, but robots were either too expensive or not performant enough.
So we built our own!
Excited to finally share something I’ve had so much fun working on:
1. Pixel: a co-designed full stack system for learning mobile manipulation.
2. Pixel-UMI: environment aware whole body manipulation without robot data.
w/ @JasonJZLiu
(1/N)
Can more pretraining improve a robot policy’s success rate, yet makes it worse for porttraining? 🤔
In our new work, we find that it can. Continued supervised pretraining of generative robot policies can improve reliability at the expense of coverage, and this tradeoff affects downstream posttraining.
Read on to see why this happens and how we can preserve coverage. 1/n 🧵
Sharing this work that I was partially involved with at @multiplylabs
We show that using contact information as feedback signal during training can help track grasps even on new, unseen objects. All demos are open loop. Excited to show this work.
Project page and video: multiplylabs.github.io/oath/
Paper: arxiv.org/abs/2610.14379
Sharing OATH: Object-Agnostic Tracking for Dexterous Hands.
We found that observing the error to reference contact points, a single residual policy can track grasps on completely unseen objects, enabling rapid physics-based retargeting.
Hi all, I have a new result I'm excited to share. It turns out that *all existing implementations* of softmax and cross-entropy (that's JAX, PyTorch, textbooks, etc) have a subtle bug --- two bugs, actually. A simple rewrite fixes both, and eliminates "slingshots" when grokking:
A million hours of what?
If we’re scaling human demonstrations for robot learning, what should each hour contain?
We built S1, our skill capture glove, alongside our robot hand to record joint movements that map to robot targets, with touch and video on the same timeline.
This is one of my favorite projects I’ve been a part of this year. It’s ostensibly about robotics, but I feel that I learned something very fundamental about how AI models should represent all modalities that (superficially!) seem like they are best suited to being represented in continuous-valued (e.g. float) vectors. TL;DR discrete parametrizations are awesome, learn geometry from data, but aren’t constrained by it, and can perform better as a result. Exactly what the bitter lesson tells us we should look for.
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].
Video model encodings with recurrent architectures (e.g. GDNs) for latent memory compression (like workspace models) seem to be scaling with the use of world models at scale in multi-task settings!
The most impressive tasks for me are the seemingly innocuous ones, like wiping a table. Purely inferring actions from states that look more or less the same can render traditional MDP formulations unsuccessful. Using latent memory states to steer policy actions seems neat to convert the POMDP to an MDP!
What to remember in the past to act—spatial, temporal or motion details?
We often assume what’s useful, but that’s hard to know in advance.
Meet IMPRINT 🧠: retain rich memory first, then learn to use it for control. Compact latents keep inference fast as history grows. 🧵 1/10