Humans learn new manipulation skills from examples and improve as they see more examples. How can we endow robots with the same ability? 🤖 🚀We introduce RoboSSM, scalable in-context imitation learning that enables robots to learn and improve at test time—robots can improve with more examples without returning to the GPU for fine-tuning. 🧵 Paper: arxiv.org/abs/2509.19658 Video: youtube.com/watch?v=YR4m21zH…

I’m in Pittsburgh for #IROS2026! 🤖 Excited to present 🔥 RoboSSM — scaling in-context imitation learning with State Space Models! Happy to chat at my poster session on Monday(9/28)! 🎤 Lightning Talk: 9:35AM @ ROOM 409/410 📍 Poster: 11:00 AM - 12:30 PM @ Kiosk E8

Humans make the dexterous task look unfairly easy. And that's what OM-1 and Omnibody Hand aims to capture.

The combination of speed and dexterity is very impressive. Congrats to the team!

.@RewardAI_ came out of stealth yesterday with OM-1, a robot foundation model trained entirely from human demonstrations. The part I find interesting is how they collect the data. Instead of teleoperating a robot, a person wears a device called the Omnibody Hand and simply does the task themselves. Pick something up, fold laundry, mix a drink, package a phone, or plug and unplug cables. The system records the human hand motion, tactile information, proximity, vision and position. That data can then be used to train OM-1 to control different robot bodies. Reward calls the idea: “One Model, One Data Interface, Any Body.” According to the company, a new task can be learned from less than 30 minutes of human demonstration data, without collecting additional robot data. The videos are worth watching. They show four robot arms packaging a phone, robots folding laundry and mixing cocktails, an arm unplugging an Ethernet cable, and early experiments with humanoids. Some of the longer tasks finish in under 30 seconds, in real time. There is also some history behind this. Reward AI builds on DexCap, work that came out of Stanford in 2024 around capturing human dexterous manipulation and transferring it to robots. @chenwang_j , C. Karen Liu and Li Fei-Fei were among the researchers involved. I think the bigger bet here is about the data layer. Robot hardware is changing quickly. If collecting training data means teleoperating every new robot body for thousands of hours, scaling gets expensive fast. Reward is betting that you can collect much of that knowledge from humans instead, then transfer it across different embodiments. That could become quite valuable if it works. For now, we only have the launch material. There is no paper or open model yet, and Reward has not published the kind of large-scale success-rate data needed to judge robustness. Still, this is one of the more interesting approaches I’ve seen to the robotics data problem recently. Source: rewardai.com/blog/OM-1/ —— Weekly robotics and AI insights. Subscribe free: 22astronauts.com

Wait, 1×speed?

So snappy, and such an underrated bit of co-design!

The speed is seriously impressive — I can only imagine how much hard work went on behind the scenes. Huge congrats to the team on the release!

Wow

daaamn, just realised this is on 1x speed smoooth

Achieving human-level speed execution is something that Astra can never achieve. Congrats to @zipengfu and the team!

Amazing!!!

Learning from human data could let robot learning scale beyond robot data collection. Cross-embodiment models make this even more interesting - shared skills across different robot bodies. A very promising direction!

Fast, smooth, with strong generalization across embodiments! Congrats to the @RewardAI_ team!

amazing co-design work behind this level of dexterity and reactivity! generalization with these skills will be the next frontier

Congrats to @zipengfu, @chenwang_j, and co on the release!

ロボットがバーテンダーの手捬きを再現しているこの映像、人間の作業データだけで学んでいる。しかも映像は等倍。 Reward AIのロボット基盤モデル「OM-1」。一言でいうと、ロボットの学習データをロボットから取るのをやめた。 ・センサー付きグローブを着けた人間が普通に作業する、そのデータだけで学習 ・テレオペも実機データもゼロ ・卓上アーム / 産業用アーム / ヒューマノイドにゼロショットで転用 ・複数ロボットの協調作業まで射程に入っている ポイントは「ゼロショット」のところ。 従来のロボット基盤モデルは、テレオペで実機を動かして集めたデータが前提だった。だからデータが機体に紐づく。機体を変えれば集め直し。OM-1はデータの入口を人間の手に置いたので、まだ存在しない機体のためのデータを今日から貯められる、とのことだ。

The speed is what you notice first. What excites me most is how we were able to transfer skills learned from human motion across different robots, without robot data or robot-specific fine-tuning. Really proud of what our team has built!

This is really good; very difficult task for a robot

The handheld gripper market just gained a strong new competitor. Congrats!