Xiaomi Opens a 38B World Model Built to Generate Robot Data
SubscribeSign in
Xiaomi Opens a 38B World Model Built to Generate Robot Data<br>Xiaomi opened a 38-billion-parameter robot world model this week and uses it to generate training data, not to control robots directly. Robbyant and Alibaba each released several robot models at once.<br>Jay Chia<br>Jul 16, 2026
Share
A world model as a data engine
Xiaomi-Robotics-U0 is a single 38B model that generates robot scenes and video, built the same way as an image or video generator. Instead of using it to control a robot, the authors use it to generate training data, and report that this raised π0.5’s success on unfamiliar real-world manipulation from 36.9% to 63.2%. They also report topping a “World Arena” leaderboard for embodied video, and that human raters preferred its generated scenes over GPT-Image-2.0’s. Those numbers are Xiaomi’s own, measured on its own evaluations, but the weights and code are open, so you can run U0 as a data generator for your own policy and check whether the gain holds outside their setup.
Thanks for reading Topic Queue! Subscribe for free to receive new posts and support my work.
Subscribe
Research
NVIDIA open-sources a benchmark that grades robot policies beyond pass-or-fail
RoboLab is an Isaac Lab benchmark that goes beyond pass-or-fail. It scores how completely and smoothly a manipulation policy completes a task, how well it holds up when the instructions, the scene, or the task length change, and exactly where it fails. It ships with RoboLab-120, a set of 120 human-curated tabletop pick-and-place tasks grouped by the skill each one tests, and the code is on GitHub. So it shows what to fix, not just which policy ranks higher.
Robbyant opens the rest of its LingBot stack
A week after open-sourcing LingBot-VLA 2.0, Robbyant released two more models, both trained on robot data from the start rather than repurposed from a model built to generate ordinary video. LingBot-Video is a Mixture-of-Experts video model trained on robot footage and rewarded for producing physically realistic clips. LingBot-World 2.0 is an interactive world model you can keep steering without it drifting over time; a distilled version runs 720p at 60fps, and it comes as a 14B model plus a 1.3B version that runs on a single GPU. With last week’s VLA, that is a full open set of models from one lab.<br>Alibaba AMAP ships an ABot model family the same week
AMAP released four models at once. ABot-C0 is a general quadruped controller trained on 16,074 motion clips, and its tracking performance continues to improve predictably as the training data grows. ABot-N1 is a navigation model that steers toward a target pixel in the camera image, an approach that works across different navigation tasks, and reports a 35-point jump in reaching named destinations. ABot-3DWorld 0 turns text, an image, or a video into a 3D scene you can move through, tied to real places on a map. ABot-AgentOS is a software layer that plans tasks and remembers across them, and it comes with a new benchmark, EmbodiedWorldBench. All of the numbers are self-reported.<br>EgoWAM asks what a world model should actually predict
EgoWAM, from Danfei Xu’s group at Georgia Tech, tests what a robot should learn to predict from everyday first-person human video. Keeping everything else the same, it compares three prediction targets: raw pixels, DINO features, or 3D motion flow. Pixels transfer poorly, DINO features improve performance on unfamiliar objects and scenes by up to 4x, and 3D flow improves performance on familiar tasks by 20 to 30%. Most world-model papers just assume a prediction target; this one measures which works. The same week, a separate paper laid out a research roadmap for world action models, the category to which EgoWAM belongs.<br>Your coding agent as a robot controller
VIA, from Dorsa Sadigh’s group at Stanford, recasts robot control as a software task: an off-the-shelf frontier agent such as Claude Code or Codex drives a manipulator through a browser-based 3D interface using screenshots and typed commands, with no robot-specific training and nothing beyond what is on screen, no exact positions of the robot or objects handed to it. On the tasks reported, the strongest agent hits 96.7% across three LIBERO-Goal tasks and 100% on a long-horizon assembly task, zero-shot. The provocation is that a capable computer-use agent can already handle some manipulation with no trained policy at all.<br>DexVerse benchmarks dexterity across arms and hands
DexVerse (UNC, HKU, Berkeley) is a simulated test suite for measuring dexterous manipulation, the fine-hand skills like grasping, using tools, and two-handed tasks. Its 100 tasks run across three robot arms and six robot hands, so the same task can be tried on different hardware, and it ships 3,180 demonstrations plus a VR tool for recording more. When the authors ran leading policies (Diffusion Policy, DP3, OpenVLA, π0.5) through it, success fell...