AgiBot Open-Sources AGIBOT WORLD 2026 Phase 4 Embodied Reasoning Dataset

"Help me get my things ready to go out." After receiving this instruction, the robot needs to first check the weather forecast, then decide which items to bring based on the weather conditions. If tomorrow is sunny, prepare sunscreen; if the weather turns rainy, bring an umbrella. For humans, this chain of judgments is almost instantaneous, but for a robot, it involves multiple stages such as task understanding, information retrieval, conditional judgment, and action execution. How can robots not only execute fixed actions according to instructions, but also autonomously judge and adjust their actions based on changes in the real environment, learning to "act according to the situation"? On October 22, AgiBot officially open-sourced the fourth installment of AGIBOT WORLD 2026, the "Physical Reasoning" dataset. Unlike data that focuses on manipulation skills, this installment shifts the task design focus from "manipulation difficulty" to "decision-making difficulty," centering on the decision-making and execution processes of the same task under different states and constraints, helping robots move from completing individual operations to autonomously handling complex tasks. Tasks in the real world often do not have a fixed operating procedure applicable to all situations. For example, to restock the beverages in a guest room minibar to 3 bottles, the robot needs to first distinguish empty bottles from full ones, count the current quantity, and then decide how many bottles to remove and restock. Facing different initial states, even with the same task goal, the actions the robot needs to take may differ. Another example is searching for toiletries across multiple drawers: the robot must not only recognize items but also remember which drawers have already been checked to avoid repeated searching. To enable robots to autonomously complete these tasks, data on "how to execute actions" alone is far from enough—there must be data on "how to make decisions based on goals and conditions." The fourth installment of the dataset is built precisely around this gap. According to the introduction, the fourth installment of the AGIBOT WORLD 2026 "Physical Reasoning" dataset constructs decision-making scenarios for the same task under different states around factors such as item quantities, occlusion relationships, position occupancy, and task progress. It covers nine core reasoning capabilities, including attribute selection, conditional judgment, operational feasibility assessment, goal search, long-horizon planning, and process memory, and sets three difficulty levels—Easy, Medium, and Hard—covering tasks ranging from basic judgments to multi-goal planning and long-horizon state maintenance. In addition to providing diverse task scenarios, the dataset also links task goals, decision-making processes, and actual actions through fine-grained annotation. The data uses task instructions to describe complete goals, and characterizes task structure through information such as task stages, subtask intervals, and action steps; it also provides detailed action descriptions and spatial annotations such as boxes, points, and lines to specify operation objects, contact positions, and motion paths, and records task execution results. These data can help researchers train robots' task understanding and reasoning capabilities, and also support coordinated training of high-level decision-making and low-level action policies. AGIBOT WORLD 2026 is an open-source dataset series launched by AgiBot for embodied intelligence research, covering five major embodied research themes, each using corresponding data collection methods and annotation systems. It is reported that this installment of the dataset will be used for training AgiBot's upcoming Agent framework and a new-generation foundation model, providing a data foundation for robots to understand tasks, memorize execution processes, and select actions.