Zhiyuan Open-Sources AGIBOT WORLD 2026 Phase 4 Dataset for Embodied Reasoning

On October 22, Zhiyuan officially open-sourced the fourth phase of its AGIBOT WORLD 2026 dataset, focused on embodied reasoning. The release shifts the design emphasis from manipulation skills to decision-making, aiming to help robots handle complex tasks by making choices based on different states and constraints, rather than following fixed action sequences.

According to Zhiyuan, the dataset builds decision-making scenarios for the same task under varied conditions, covering factors such as object counts, occlusion, position occupancy, and task progress. It includes nine core reasoning capabilities, including attribute selection, condition evaluation, feasibility assessment, target search, long-horizon planning, and process memory, with three difficulty levels: Easy, Medium, and Hard.

The dataset uses task instructions to describe complete goals and annotates task phases, sub-task intervals, and action steps. It also provides detailed action descriptions and spatial annotations such as boxes, points, and lines to identify target objects, contact positions, and motion paths, along with task execution results. The data is designed to support training for both task understanding and reasoning, as well as coordinated learning of high-level decisions and low-level action policies.

AGIBOT WORLD 2026 is an open-source dataset series from Zhiyuan for embodied intelligence research. The company said the new dataset will be used to train its upcoming Agent framework and next-generation foundation model, providing data for robots to understand tasks, remember execution processes, and select actions.