AGIBOT releases real‑world dataset

A robot dataset is only useful if it shows what robots actually do, and AGIBOT has now open-sourced a new one built from real-world runs on its G2 platform. (agibot.com) AGIBOT said AGIBOT WORLD 2026 was released in April 2026 as a heterogeneous dataset for five research tracks in embodied intelligence, the field that trains machines to perceive and act in physical spaces. The company’s site says the data comes from “100% real-world environments,” including homes, commercial spaces, and other general-purpose settings. (agibot.com) (agibot-world.com) The dataset was collected on AGIBOT’s G2 robot with synchronized multimodal signals, including red-green-blue-depth images, tactile data, LiDAR point clouds, inertial measurement unit data, and full-body joint states. AGIBOT said it also released matching 1:1 digital-twin simulation data for the same scenes. (agibot.com) (agibot-world.com) Imitation learning is the robotics version of learning by watching an expert, and AGIBOT says this dataset is built for that use case. Its annotation scheme includes high-level instructions, segment-level task descriptions, step-level action sequences, atomic skill labels such as “pull” and “place,” 2D boxes on target objects, and labeled error-recovery trajectories. (agibot-world.com) The collection method matters because many robot datasets rely on repeated, scripted demos in controlled settings. AGIBOT said its operators used a “free-form” approach, changing steps based on real conditions so the data captures variation in object types, starting positions, and task order. (agibot.com) (agibot-world.com) AGIBOT also said the system records force-controlled interactions and uses first-person teleoperation beyond the operator’s direct line of sight, so the data includes contact and force feedback rather than just camera frames and arm motions. That is aimed at tasks where a robot has to feel its way through a pull, press, or grasp, not just see it. (agibot.com) The release extends an AGIBOT World program that already included Alpha and Beta datasets, a foundation model called GO-1, and a benchmark ecosystem. The OpenDriveLab repository for the earlier AgiBot World project lists a Beta release with 1,003,672 trajectories, an Alpha subset with 92,214 trajectories, and more than 100 replicated real-life scenarios across five domains. (github.com) The new dataset is also tied to AGIBOT’s competition push around the IEEE International Conference on Robotics and Automation in 2026. AGIBOT said its AGIBOT World Challenge uses the G2 robot, Genie Sim 3.0, and two tracks focused on “Reasoning to Action” and “World Model,” with registration announced in February and track details updated in March. (agibot.com 1) (agibot.com 2) (agibot-world.com) The public hosting is live on Hugging Face, where the dataset card lists robotics tasks, image and text modalities, English language labels, and a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 license. The preview page shows the train split, though the full viewer was not available when accessed because the job manager had crashed. (huggingface.co) AGIBOT is pitching AGIBOT WORLD 2026 as a way to move robot training out of tidy lab demos and into messier everyday environments. The test now is whether outside labs use the open release to build systems that transfer from simulation and teleoperation into reliable physical behavior. (agibot.com) (agibot-world.com)