Xiaomi trained a robot foundation model on 100,000 hours of real-world manipulation data, and the scaling laws show no signs of stopping

Xiaomi Just Proved Scaling Laws Work for Physical Robots

The core thesis behind the large language model explosion was brutally simple: give a model more data and more compute, and it gets better, reliably. That empirical scaling law turned GPT-class models from lab demos into infrastructure in roughly two years. Xiaomi just showed the same curve exists for physical robots.

Xiaomi-Robotics-1 is a vision-language-action model pre-trained on over 100,000 hours of real-world manipulation trajectories, collected across more than 1,700 scenarios spanning homes, businesses, and industrial sites 1. The training recipe mirrors the LLM playbook: pre-train on massive, broad data, then post-train on task-specific examples to align the model with real-world use. Code and model checkpoints are being released openly 2.

The finding robotics has been waiting for: the scaling curve is clean and shows zero saturation. As pre-training data and model size grow, real-robot success rates in unseen environments improve steadily and predictably 1. That relationship between scale and capability is the engine that drove the language model boom, and it has been conspicuously absent from robotics, where high-quality training data has been the binding constraint 1.

The structural significance is in the split. Pre-training, the expensive stage, happens once on general data and produces a broadly capable foundation. Post-training, the cheaper stage, specializes that foundation for specific robot hardware and task instructions from a fraction of the data. That division of labor is exactly what made the LLM stack economically viable: one massive training run, then thousands of cheap adaptations. Xiaomi has now replicated it for the physical world 1.

Breaking the data constraint required a pipeline built for volume no human labeling team could match. Xiaomi collected 100,000 hours of manipulation trajectories using embodiment-free UMI devices across household, commercial, industrial, and outdoor environments 1. An automated annotation pipeline powered by a vision-language model splits long videos into clips and labels each one with descriptions of how objects and grippers change state, producing the language conditioning the robot policy learns from 1.

Post-training layers in real robot data: over 7,200 hours collected inside actual homes, covering tasks like tidying a sofa, sorting a shoe cabinet, and putting away kitchenware 1. This stage maps the general action-generation ability onto physical robot hardware and shifts the model from describing scene transitions to following natural-language instructions directly 1.

When learning brand-new tasks like phone packing, printer refilling, laundry loading, and box packing, Xiaomi-Robotics-1 reaches 75% success from an average of under 10 hours of demonstrations per task, compared to 40% for the pi-0.5 baseline at the same data budget 1. Raising the budget to under 40 hours per task pushes overall success to 85% 1.

On simulation benchmarks, the model sets state-of-the-art results across four mainstream suites: RoboCasa, RoboCasa365, VLABench, and RoboDojo 1. The biggest gains land on the benchmarks that stress-test generalization, with a 58.3% relative improvement over the previous best on RoboDojo and 23.2% on RoboCasa365 1.

The absence of saturation matters more than any single benchmark score. If the curve keeps climbing with more data and larger models, the limiting factor shifts from algorithmic breakthroughs to data collection and compute budget. That is the exact transition that turned language models into a capital-intensive arms race, and it is now on the table for robotics 1.

If robot foundation models follow the same scaling trajectory that language models rode from 2022 through 2024, the distance between what robots can do today and what they will do next year compresses fast. The company proving this is not OpenAI, Google DeepMind, or a Western robotics startup. It is Xiaomi, a Chinese consumer electronics company, releasing the code and weights for anyone to build on 2. The pre-train-then-post-train paradigm has crossed from text into the physical world. The question is how quickly the rest of the field catches up, and whether the robotics scaling race will look anything like the one that remade the AI landscape.