Poke-Aurora

Poke-Aurora is PokeBot's first world-action model, publicly introduced October 29–30, 2026 (Beijing) together with RoboAPI. It takes wrist-camera observations plus a natural-language instruction and predicts the next robot action chunk in near real time—aimed at zero-shot manipulation without per-task fine-tuning.

Product: pokebot.com/ai · Docs: developers.pokebot.com/docs/roboapi · X: @PokeBot_Global

Key points

1. Named model: Poke-Aurora (world-action / 世界动作模型), distinct from the RoboAPI serving layer.
2. Vendor claim: 74.2% task success across ~1,000 unseen objects and ~10,000 unseen tasks; says it beats pi0.5 and GPT-6 Astra on their internal suite (vendor numbers).
3. Tech teaser: Modulated Residual Coupling (MRC) between video prediction and action heads; UMI-style wrist fisheye + end-effector pose control expected.
4. Access path: free non-commercial RoboAPI technical preview; email registration opened Oct 30 BJT.
5. Traction: official @PokeBot_Global launch (~17k impressions) plus robotics Twitter/Chinese tech coverage the same window.

Why it matters

Embodied-AI searchers will look up Poke-Aurora specifically as the model behind RoboAPI—clear proper noun with a public try-it API.

How to try it

1. Register at platform.pokebot.com for RoboAPI preview access.
2. Connect a compatible arm (UR / Franka-class with wrist fisheye + IK) per the RoboAPI docs.
3. Send observation + plain-language instruction; execute predicted action chunks carefully in a safe cell.

Preview is non-commercial; treat vendor success rates as unverified until third-party reports land.