Xiaoshen Han

Excited to share B-spline Policy ⚡, a project that Haoyu and I co-led. We rethink action representations to enable faster robot manipulation!

Introducing Waddle Labs: Claude Code for robots. Connect our API to your robot and enter a prompt, then our agents write code to achieve the task in 20 minutes. @yiding_song @theWaddleLabs

Today, we’re launching Tau’s humanoid cleaning service in San Francisco at $30 per hour. Access is initially invite-only as we scale operations. If you don’t have an invite yet, join the waitlist at tau-robotics.com. All footage is shown at 1× speed. Each humanoid is jointly controlled by a human operator and AI.

Spline policy arxiv.org/pdf/2606.07386 B-Spline policy arxiv.org/pdf/2607.09648 Both showing spline action parameterization helps manipulation learning, with the first one having more equations.

Releasing the model weights and technical report of Kimi K3. Kimi K3 is our most capable model: a 2.8T MoE model with native visual understanding and a 1M-token context window. New model architecture: 2.5x the intelligence per unit of compute, not just more params. Alongside Kimi K3, we're opening up more of the stack behind it — high-performance attention kernels, MoE communication library, and infrastructure for running agent environments at scale. Model weights: huggingface.co/moonshotai/Kim… Tech report: github.com/MoonshotAI/Kim… Tech blog: kimi.com/blog/kimi-k3

We released τ₀-VLA, a robot foundation model built for long-horizon real-world manipulation. Robots can now clean rooms, cook meals, make drinks, collect laundry, and organize objects in autonomous episodes lasting up to 12 minutes. 🧵

Excited that OAT (ordered-action-tokenization.github.io) was selected as one of the featured research projects at the YCML Research Symposium hosted by @ycombinator! Unfortunately, I won’t be able to attend due to an unexpected U.S. re-entry issue. My co-author Jiawei (@WinstonGu_) will be presenting our work on our behalf. If you’re attending #yc AI Startup School, stop by our poster this Sunday, 2–3 PM PT and chat with Jiawei about action tokenization for robot learning!

Paper: arxiv.org/abs/2607.09648 Blog: b-spline-policy.github.io GitHub: github.com/B-spline-polic… Colab: colab.research.google.com/drive/1MzbUGoL…

B-spline Policy has real limits. Cheap arms hit a wall: on Speed Stacking, 4X dropped to 0/20 because the controller couldn't track that fast. The fitting tolerance still needs per-task tuning. Still, continuous action representations look like a real path to faster manipulation.

B-spline Policy's results hold up across real tasks. On Cube Picking, adding it to Diffusion Policy at 4X reached 20/20 in 2.45s. The baseline reached 19/20 in 6.48s. On Table Cleaning it beat rival DemoSpeedup 14/20 to 3/20 at roughly 4x the speed. Cup stacking doubled to 16/20

Instead of discrete waypoints, B-spline Policy represents a robot trajectory as a continuous curve: a compact set of knots and control points. That curve can be sampled at any frequency, and you can speed the robot up just by traversing it faster. No retraining per target speed.

State-of-the-art robots take close to a minute to fold a T-shirt while humans do it in about 10 seconds. Success rates have climbed for years, but speed has lagged behind. A new method called B-spline Policy by @xshenhan and @Haoyu_Xiong_ goes after speed directly. 🧵👇

@cybernetic_lab @Haoyu_Xiong_ Thank you!

We just hit a weird milestone: our model became more reliable than your average home WiFi. Just like everybody else, we thought cloud inference was the obvious choice. Yet 2 days into the ACT-2 eval, our mind completely changed. If our hero @ArpitKalla didn’t cook, this video wouldn’t exist 🧵

Introducing ACT-2 Preview, the world’s first robotics model that works in your home. 99% success rate, fully autonomous in unseen homes. Zero data from you.

@HammadTime thank you!🥰

clever

Introducing ACT-2 Preview The first robotics model to unify broad generalization with high reliability. A single fine-tuning example can teach Memo a new behavior that generalizes. Zero shot, real unseen homes, 99% success rate.

Can robot foundation models scale? Xiaomi-Robotics-1 explores this question with over 100,000 hours of real-world manipulation data. Pre-trained on large-scale real-world trajectories and post-trained with cross-embodiment robot data, Xiaomi-Robotics-1 demonstrates consistent scaling across data and model size, strong generalization in unseen environments, and efficient adaptation to new tasks. 🔗 robotics.xiaomi.com/xiaomi-robotic… #Robotics #EmbodiedAI #FoundationModels #RobotLearning #XiaomiAI #XiaomiRobotics

@IlirAliu_ thank you～🥰

Posted by Haoyu Xiong: x.com/Haoyu_Xiong_/s… Thread (technical explanation + experiments) By Xiaoshen Han (co-lead): x.com/xshenhan/statu…

This policy enables fast manipulation on low-cost robot arms: B-spline Policy (BSP), which represents robot actions as continuous B-spline curves instead of discrete fixed-rate chunks to enable faster, smoother manipulation on low-cost arms. Side-by-side real-time videos compare BSP against diffusion and regression policies, showing it completes tasks like dish handling, cup stacking, and table setting significantly quicker by adapting speeds for free space versus contact phases. @Haoyu_Xiong_, an MIT CSAIL PhD student, co-led this recent arXiv paper with @xshenhan and collaborators from Harvard and UT Austin. The work looks not only at success rates but also at how quickly a robot can actually complete a task. 📌 Paper arxiv.org/abs/2607.09648 ——- Weekly robotics and AI insights. Subscribe free: 22astronauts.com