Event

AgiBot Open-Sources Embodied Reasoning Dataset

1 reports · 1 sources

On October 22, AgiBot open-sourced the fourth installment of AGIBOT WORLD 2026, the "Embodied Reasoning" dataset, which it describes as the first real-world Physical Agent dataset. The dataset covers nine categories of logical reasoning scenarios, constructing multiple real-world states and execution branches around the same task, while providing both high-level question-answer pairs and low-level action trajectories for coordinated training from task decision-making to real-machine execution. Tasks are graded as Easy, Medium, and Hard: basic tasks test single-item judgments such as attributes and shape matching; medium tasks add quantities and preconditions; high-difficulty tasks layer multi-subgoal planning, long-horizon state maintenance, and contextual constraints. AgiBot says the dataset has already been used to train its upcoming Agent framework and its new-generation foundation model. The above claims of "first" and of the training use are AgiBot's own statements.