The Pulse
PokeBot Opens RoboAPI to Test Its Poke-Aurora Robot Model
PokeBot opened RoboAPI, an API that connects compatible robot arms to its Poke-Aurora world-action model, for a free, noncommercial technical preview. The company reports a 74.2% success rate on its own evaluation of nearly 10,000 tasks, wh
AI.info Team · Written with AI · fact-checked ·
“We also need to make it easier for developers to put embodied AI to the test in the real world,”
Huazhe Xu, founder of PokeBot and an assistant professor at Tsinghua University’s Institute for Interdisciplinary Information Sciences
PokeBot opens access to Poke-Aurora
PokeBot opened RoboAPI on October 30, giving developers a way to connect compatible robot arms to Poke-Aurora, the company’s world-action model. The API takes a robot’s observations and state along with a user’s instruction, then predicts a chunk of actions for the machine to execute. The company says developers can try the service without collecting task-specific data or training a new model for each job.
The launch follows PokeBot’s announcement on October 29. RoboAPI is in a technical preview, offered free for noncommercial use; the company says it does not provide products or services to for-profit third parties. Registration opened October 30 through PokeBot’s platform.
A wrist camera is central to the setup
Poke-Aurora combines predictions about future visual observations with robot-action predictions. PokeBot says it added a method called Modulated Residual Coupling to encourage the model’s video-prediction and action-prediction components to inform one another. The company says more technical details will appear in a technical report.
For robot control, PokeBot aligned the model around a wrist-mounted fisheye camera and end-effector pose control. The company says this setup transferred from a UR7e arm to a YAM Standard configuration, and that changing between two camera brands had almost no effect on performance in its tests. It says the approach still depends on the arm having a sufficiently capable inverse-kinematics solver.
The 74.2% result comes from PokeBot’s own test
PokeBot evaluated Poke-Aurora on nearly 1,000 distinct manipulable objects and nearly 10,000 instructions, grouped into object transfer and organization, pose and shape alignment, and environment interaction. The company reports an overall task success rate of 74.2%, compared with GPT-6 Astra, π0.5 and MolmoAct2. Those figures come from a benchmark PokeBot built; the company notes that robot setups and task definitions complicate comparisons across models.
Results varied by task. Poke-Aurora scored 86.0% on transfer and organization, 54.1% on pose and shape alignment, and 87.7% on environment interaction. Its weakest category was aligning poses and shapes, despite outperforming the listed comparison models in each category. PokeBot cautions that users may not reproduce its benchmark results in their own settings.
External tests will expose the gaps
PokeBot also describes internal trials across dozens of settings and a private beta in which participants used their own robots and everyday objects. Examples on the launch page include opening an air fryer, placing bread into a toaster, stacking cups and standing a bottle upright. The company presents these as task demonstrations, not as a controlled independent evaluation.
The setup has practical limits. PokeBot warns that an object held by the arm can block the single camera’s view, reducing performance; it also says shorter, precise English instructions generally work better than long tasks. The company recommends setting safety boundaries around the arm or using compliant control, and acknowledges that RoboAPI can make mistakes even though its safety gate rejects instructions it identifies as potentially harmful.
Opening the API gives outside developers a way to test Poke-Aurora on physical hardware, beyond the company’s own benchmark and demonstrations. Whether its reported transfer between robot setups holds up across users’ equipment and tasks is the test RoboAPI now makes possible.