X-Humanoid unveils roadmap for unified embodied model Pelican-Unify

Beijing's X-Humanoid center used an October 17 keynote at ROSCon China 2026 in Shanghai to outline a unified embodied model and Pelican-Unify, which links scene understanding, task reasoning, future video and action generation. No release date was given; Pelican-Sim raised policy success on RoboTwin from 70% to 93% with 500 generated trajectories added.

• Four generalization levels: background, compositional skills, cross combinations and out-of-domain tasks
• Pelican-Sim 1.0 was trained on about one million trajectories and predicts scene changes
• Adding 500 generated trajectories lifted RoboTwin success from 70.0% to 93.0%
• Pelican-VLA 0.5 has been on GitHub since July; Pelican-Unify release date unannounced