PokeBot

We had a lot of fun testing out RoboAPI, the set up is very easy if you want to try it out yourself! :)

Really enjoy playing with Poke’s RoboAPI in our lab. The whole system is plug-and-play. I’ll definitely stay tuned and look forward to their future updates. teapawt.org/PokeBot_Global/status/…

Super Cool! Look forward to more sharing about it's eval. Zeroshot on unseen object based on PokeBot WAM. UMI-style post-training and leverage wrist cam and standardized EEF control to adapt to different arms, If you have Franka / UR arm, have a try!

PokeBot has launched RoboAPI, a public API that lets robot arms perform new tasks without collecting data or training a dedicated model. Users can connect a supported arm with a wrist-mounted camera, give it a task such as "clean the table", and have RoboAPI send back its next moves in near real time. #PokeBot

We had a lot of fun testing out RoboAPI, the set up is very easy if you want to try it out yourself! :)

RoboAPI: Zero-Shot Robot Control Through an API @PokeBot_Global's official account (破壳机器人), has introduced RoboAPI on Zhihu: an API designed to let robots follow new instructions without task-specific data collection or retraining. Send robot observations, state and an instruction; receive predicted actions in near real time. Connect world prediction to action RoboAPI runs on Poke-Aurora, a world-action model that jointly predicts future video and robot actions. Its key mechanism, Modulated Residual Coupling (MRC), connects the video and action modules while preserving their native computation paths. The team reports better alignment between their learning signals, helping knowledge of physical interactions translate into more reliable actions. A shared interface across different robots Post-training uses a UMI-inspired setup: a wrist-mounted fisheye camera and end-effector pose control. This reduces dependence on a robot’s particular joint layout. The team reports transfer from UR7e to an unseen YAM Standard configuration. Reliable inverse kinematics is required, and performance varies across hardware. Generalization users can test themselves The team’s internal evaluation spans nearly 1,000 everyday object instances and nearly 10,000 instructions, alongside trials by external users with their own robots. Occlusion remains a limitation, and precise English instructions or shorter subtasks generally work better. Access is free during the technical preview, subject to non-commercial-use restrictions.

Teaching a robot a new task has traditionally meant collecting data, training models and repeating the process until it works. PokeBot wants to make that much simpler. Its new RoboAPI lets developers give robots instructions in everyday language and try Poke-Aurora on supported hardware, without having to train a model from scratch. Making advanced robotics easier to experiment with could open the door to a lot more innovation.

A lot of embodied AI still asks you to trust a lab video. @PokeBot_Global is opening RoboAPI instead. Non commercial under PokeBot. Connect supported hardware, give natural language instructions, see how Poke Aurora holds up on your own tasks without fine tuning first. They show the footage at normal speed. pokebot.com/ai Early access: platform.pokebot.com API docs: developers.pokebot.com/docs/…

This is truly impressive work! I used RoboAPI to reproduce a zero-shot task, and it worked on the very first try. Even more impressively, the robot knows when something goes wrong and can retry on its own. Huge thanks to @HarryXu12 and his team @PokeBot_Global for their incredible work!

Is robotics having its API moment? Physical Intelligence announced its Pi API beta earlier. Now PokeBot has launched RoboAPI, giving developers access to its Poke-Aurora world-action model. Developers can tell their robots what to do in plain language, without first collecting task-specific training data or fine-tuning the model. PokeBot has already opened it to outside testers using their own robots and environments. Turns out there’s an even bigger surprise at the end: PokeBot slipped in a humanoid robot teaser… 👀

Robot intelligence is moving from the lab bench to the cloud. PokeBot released RoboAPI: an out-of-the-box embodied API that lets you run its Poke-Aurora World-Action Model on supported hardware with just natural-language prompts. It's capable of zero-shot generalization at ~1x human speed, and it's openly verifiable: test it in your own environment, on your own tasks. RoboAPI is a non-commercial product under PokeBot. Docs: developers.pokebot.com/docs/…

Physical AI models are moving toward an API-driven ecosystem. Physical Intelligence has opened its Pi API beta. Dream Labs has launched a public API for Dream-1. Generalist is expanding access to GEN-1.5 through a waitlist. Now PokeBot Robotics is joining this shift with RoboAPI, a non-commercial product that gives developers access to Poke-Aurora, its first World-Action Model, on supported hardware. Similar to what APIs did for LLMs, this could make robot intelligence more accessible to developers. Rather than building and training a model from scratch, developers can start with existing capabilities and focus on adapting them to real-world applications. This could also accelerate robot deployment. More system integrators, vertical robotics companies and industry solution providers could use foundation models as a starting point, focusing on validating robot behavior in customer environments, building reliable workflows and scaling deployments. The deeper shift works in both directions. For foundation model companies like PokeBot, an API ecosystem creates more opportunities to test models across tasks and environments, expose failure cases and accelerate iteration. Meanwhile, developers and deployment teams bring general capabilities into specific customer workflows, where they encounter real-world constraints and discover what it takes to make robots productive. Some of the most capable and productive Physical AI systems may emerge from this feedback loop between foundation models and real-world deployment. The API is the access layer. Real-world deployment is where models prove their value, and where the next generation of Physical AI may take shape.

.@PokeBot_Global has opened RoboAPI to developers, starting with its Poke-Aurora model. It only runs on supported hardware. Poke-Aurora is what PokeBot calls a World-Action Model (WAM). You give it an instruction in plain language plus the robot’s current state, and it sends back robot actions in near real time. PokeBot says the footage in its demos, including the video below, plays at 1× speed. Right now, getting a robot to do one new task usually means a loop: collect demonstrations, train, test, then collect more. RoboAPI is meant to skip that loop. You follow the hardware setup guide and then prompt the robot. Because anyone with a supported robot can run it, the model can be tested outside PokeBot’s own lab. PokeBot was founded in April 2026 out of Tsinghua University’s Institute for Interdisciplinary Information Sciences (IIIS) and the Shanghai Qi Zhi Institute. Its founding team is Huazhe Xu, an assistant professor at Tsinghua IIIS whose work won the CoRL 2023 Best System Paper award; CTO Zhengrong Xue; and Chief Scientist Zhecheng Yuan. Its investors include Shunwei Capital, Matrix Partners China, Xiaomi and Baidu Ventures. RoboAPI is a non-commercial product from PokeBot, and it’s free to try. Tech blog: pokebot.com/ai Apply for early access: platform.pokebot.com API docs: developers.pokebot.com/docs/… #PokeBot