AGIBOT releases large real‑world dataset
AGIBOT published AGIBOT WORLD 2026, an open dataset of real‑world home and commercial robot runs covering manipulation, long‑horizon tasks, dual‑arm actions and human‑robot collaboration, and it includes failures and recoveries plus sim twins. The release is aimed at researchers and builders who need physical interaction data rather than synthetic traces, and it highlights the value of failure cases for robust policy learning. Public real‑world datasets like this can shorten development cycles by providing tangible training and evaluation material for embodied AI. (x.com)
Robots do not fail in the neat way software fails. A chatbot gives you a wrong sentence; a home robot misses the mug handle, bumps the counter, and has to try again with force and timing that change every second. (agibot-world.com) That is why robot builders care so much about data from the physical world. A model can learn from internet text or simulated images, but opening a cabinet, folding fabric, or handing a box to a person depends on friction, weight, contact, and awkward starting positions that are hard to fake. (agibot-world.com) AGIBOT’s new release is a dataset built for exactly that problem. The company says AGIBOT WORLD 2026 is collected from 100 percent real-world environments, including homes, commercial spaces, and other general-purpose settings, rather than only from simulation. (agibot-world.com; therobotreport.com) The collection method is less like filming the same factory demo 1,000 times and more like letting operators improvise inside a real task. AGIBOT says teleoperators used a free-form collection mode, so the runs vary across object types, starting positions, and operation sequences. (agibot-world.com; humanoidroboticstechnology.com) The first phase focuses on imitation learning, which is the robot version of learning by watching an expert. AGIBOT says each episode includes high-level instructions, step-by-step action breakdowns, atomic skill labels such as pull and place, and two-dimensional boxes marking target objects. (agibot-world.com) The unusual part is that the mistakes stay in. AGIBOT says error-recovery trajectories are retained and annotated, which means the dataset includes not just successful grabs and placements, but also the missed attempts and the correction steps that follow. (agibot-world.com) That matters because real work is mostly recovery. A warehouse robot that can only succeed from a perfect starting pose is like a self-driving car that only works on an empty road; the valuable skill is noticing the drift and getting back on track. (agibot-world.com) AGIBOT also paired the physical runs with one-to-one digital twins in simulation. The company says it built simulation scenes at the same scale as the real environments and released that simulation data alongside the real-world data, which gives researchers a way to test ideas cheaply before pushing them back onto hardware. (agibot-world.com; humanoidroboticstechnology.com) This release sits on top of a bigger AGIBOT data effort that already claimed scale most labs cannot match. The AgiBot World Colosseo paper described more than 1 million trajectories across 217 tasks in five deployment scenarios, and the authors reported that policies pretrained on that data outperformed ones trained on Open X-Embodiment by an average of 30 percent. (arxiv.org; github.com) The new 2026 dataset is also being packaged where researchers actually fetch training data. A public Hugging Face dataset card is already live under `agibot-world/AgiBotWorld2026`, tagged for robotics, imitation learning, real-world data, and dual-arm tasks. (huggingface.co) The bigger bet is that robotics may need its own version of the giant public corpora that accelerated computer vision and language models. AGIBOT is explicitly framing this release as infrastructure, not a one-off demo, and its 2026 challenge at the Institute of Electrical and Electronics Engineers International Conference on Robotics and Automation is built around the same stack of data, simulation, and real robots. (agibot-world.com; agibot.com)
