AGIBOT opens real‑world dataset

Robots learn physical tasks from examples the way a trainee copies a skilled worker, and AGIBOT has now opened a real-world dataset built for that job. (agibot.com) AGIBOT said on April 7, 2026 that it released AGIBOT WORLD 2026 as an open-source dataset for embodied intelligence, the field that tries to make machines perceive, move, and manipulate objects in the physical world. The company said the release is organized around five research pathways, with the first phase focused on imitation learning. (agibot.com) The project site says the data was collected in “100% real-world environments,” including homes, commercial spaces, and other general-purpose settings, using AGIBOT’s G2 robot platform. The dataset is also mirrored on Hugging Face under the name “AgiBotWorld2026,” where the card lists robotics as the task category and a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 license. (agibot-world.com) (huggingface.co) Imitation learning is the first release because it trains a robot from demonstrations instead of hand-written rules, and AGIBOT says this phase includes high-level instructions, segment-level task descriptions, step-by-step action sequences, atomic-skill labels, and 2D boxes around target objects. The site also says it keeps and labels error-recovery trajectories, which are examples of how a task gets back on track after a mistake. (agibot-world.com) AGIBOT says the collection method differs from scripted robot demos because teleoperators changed steps in real time as scenes changed, rather than repeating a fixed sequence. The company says that “free-form” setup is meant to widen variation across object types, starting positions, and action order. (agibot.com) The release also pairs physical data with a matching simulation copy of each scene, which AGIBOT describes as a 1:1 digital twin. The Hugging Face readme says the simulation side is being open-sourced through the Genie Sim project alongside the real-world data. (agibot.com) (huggingface.co) AGIBOT says the robot records more than camera video: the pipeline includes red-green-blue-depth images, tactile signals, LiDAR point clouds, inertial measurement unit data, and full-body joint states. The company also says it used force-controlled collection so the dataset captures contact and pressure, not just where the robot moved. (agibot.com) The timing lines up with a larger push around benchmarks and competitions for physical artificial intelligence. The IEEE International Conference on Robotics and Automation lists an AgiBot World Challenge for 2026 with tracks in world models, vision-language-action systems, and whole-body control. (2026.ieee-icra.org) That leaves researchers with a concrete next step: download the imitation-learning phase now, test it against AGIBOT’s challenge tracks, and see whether robots trained on messy household and commercial scenes behave better outside the lab. (huggingface.co) (2026.ieee-icra.org)