PokeBot Opens RoboAPI to Stress Test World-Action Model
AI-generated image · Bay Street Wire
The new interface allows developers to deploy Poke-Aurora on physical hardware to evaluate generalization across unseen tasks and environments.
PokeBot has launched RoboAPI, a new interface designed to let developers test the company's Poke-Aurora world-action model on compatible robotic systems. According to reporting from the Financial Post, the tool is intended to lower barriers for embodied AI evaluation by removing the need for developers to perform extensive data collection or train new models from scratch for every specific application.
For hardware developers, the interface serves as a bridge to navigate the "sim-to-real" gap. RoboAPI allows Poke-Aurora to process physical observations and instructions to generate predicted robot actions in real-world environments. This enables a critical test of whether the model's generalized weights can function in unstructured settings that differ from its original training data.
In a large-scale evaluation involving nearly 10,000 unseen tasks and almost 1,000 unseen objects, the Financial Post reports that Poke-Aurora achieved an average task success rate of 74.2%. This performance surpassed GPT-6 Astra by 16.4% and pi0.5 by 51.3%.