The Beijing Innovation Center of Humanoid Robotics, known as X-Humanoid, used a keynote at ROSCon China 2026 in Shanghai on October 17 to lay out how it plans to build what it calls a unified embodied model. Zou Shilong, the center's world model lead, gave the talk, titled Exploring Unified Foundation Models for Embodied Intelligence. The pieces he described are mostly already public. What is new is the full roadmap and a model called Pelican-Unify.
The center splits generalization into four levels. Background generalization means still completing a task after the table, lighting or backdrop changes. Compositional generalization means chaining atomic skills such as opening, grasping, rotating and hanging into new long-horizon tasks. Cross generalization covers new combinations of scenes, objects, actions, viewpoints and robot bodies. Out-of-domain generalization means reasoning and adapting when tasks and conditions fall outside the training data.
To get there, it proposes four layers of data and capability. The first three are data without a robot body, such as first-person video and handheld UMI capture; simulated or generated data; and failure-recovery data that records deviations, corrections and recovery during real or simulated runs. Together they are meant to build a robust world model simulator. The fourth layer uses inference inside that world model, plus policy iteration, so the model keeps learning from feedback.
Three projects support the plan. Pelican-VLA, whose 0.5 code has been on GitHub since July, routes perception to action through 32 bottleneck slots to keep only task-relevant features. Pelican-Sim 1.0, described in an arXiv paper on September 10, predicts how a scene will change from the current image and an action trajectory across many robot types. It was trained on about one million trajectories. Pelican-Unify, the newest piece, uses shared task representations and joint training to link scene understanding, task reasoning, future video and action generation. The center did not say when Pelican-Unify will be released.
The Pelican-Sim paper reports that adding 500 generated trajectories to 50 original demonstrations per task raised a policy's average success rate on the RoboTwin benchmark from 70.0% to 93.0%. The center also described analyzing representations, intervening on attention pathways and running behavior tests to see which visual and language signals its models actually rely on.
Next, X-Humanoid will pursue two tracks. One is the unified model, which aims to adapt across tasks, scenes and robot bodies. The other is the world model simulator, with more interaction data and better long-horizon and failure-recovery simulation. It also plans a loop that links in-model rollouts, policy updates, real-world validation and the data that flows back.
SEE ALSO: UBTECH Details Thinker-WM, a World Model That Unifies Video, Depth and Robot Actions · Xiaomi Open-Sources Robotics-U0 Embodied World Model and Training Stack