# Xiaomi-Robotics-0: An Open-Sourced Vision-Language-Action Model with Real-Time Execution Source: https://www.youtube.com/watch?v=hmLkQ6Ob6Jo Recap page: https://rapidrecap.app/video/hmLkQ6Ob6Jo Generated: 2026-02-18T13:34:07.834+00:00 --- ## Quick Overview Xiaomi Robotics Zero introduced an open-source Vision-Language-Action (VLA) model that achieves real-time execution by decoupling cognitive prediction from motor control, resulting in significantly faster and smoother action generation compared to previous models, even when tasked with complex actions like towel folding. **Key Points:** - Xiaomi Robotics Zero developed an open-source VLA model that separates cognitive prediction (what to do) from motor control (how to do it). - The model successfully solved the 'Stop-and-Go' problem by achieving continuous, smooth paths for actions like towel folding, unlike models that rely on discrete point-to-point planning. - The model was trained on a massive 200 million time-step dataset, mixing open-source sets (like Droid and MoMa) with proprietary data. - When tested on the Towel Folding task, the new model achieved a 98.7% success rate, significantly outperforming baseline models like POPE (Object-level Pose Estimation) and PIO-0. - The VLA model demonstrated superior performance by maintaining a 1.0 piece-per-minute baseline for the task, compared to 0.5 pieces per minute for the baseline. - The researchers found that the best predictor for the next action was the model's own previous action, suggesting that models often get stuck in repetitive loops if not properly constrained. - The new architecture forces the model to ground its predictions in reality by paying attention to the actual current state, preventing it from hallucinating or drifting. ![Screenshot at 01:27: The introduction of the Xiaomi Robotics Zero model is highlighted, emphasizing its goal to address persistent frustrations in modern robotics by integrating real-time execution.](https://ss.rapidrecap.app/screens/hmLkQ6Ob6Jo/00-01-27.jpg) **Context:** The video discusses a research paper from Xiaomi Robotics Zero introducing a novel open-source Vision-Language-Action (VLA) model designed to improve real-time execution in robotics. The central challenge addressed is the 'Stop-and-Go' problem, where robots hesitate or stutter when planning the next step of a continuous movement. The paper proposes a fundamental architectural change to decouple the high-level cognitive planning from the low-level motor control to ensure smoother, faster execution. ## Detailed Analysis The paper from Xiaomi Robotics Zero presents an open-source VLA model with real-time execution capabilities, directly addressing the persistent issue of choppy movement in modern robotics, often referred to as the 'Stop-and-Go' problem. The core innovation is decoupling the cognitive prediction (what to do) from the motor control (how to move). This allows the model to generate a continuous, smooth path for actions, rather than relying on discrete, stop-start movements. The model, which is 16 layers deep, was trained on a massive dataset of 200 million time-steps, combining open-source data like Droid and MoMa with proprietary data. The researchers highlight that the best predictor for the next action is often the model's own previous action, leading to repetitive loops if not constrained. To combat this, the model uses a Lambda-shaped attention mask to force grounding in the actual current state, preventing hallucinations based purely on its own internal predictions. When tested on the Towel Folding task, the model achieved a 98.7% success rate, performing significantly better than baseline models like POPE and PIO-0, which struggled with the task. The synchronous model showed a higher success rate than models trained solely on general knowledge, demonstrating the value of grounding the model's predictions in reality, even if it requires slightly more latency (80 milliseconds for the GPU). ### Paper Introduction - Xiaomi Robotics Zero's open-source VLA model addresses the 'Stop-and-Go' problem in modern robotics - Introduces decoupling cognitive prediction from motor control for real-time execution - Model runs on a single Nvidia GeForce RTX 4090. ### Methodology and Training - Model trained on 200 million time-steps, mixing open-source sets (Droid, MoMa) with in-house data - Uses a Lambda-shaped attention mask to enforce reality grounding and prevent hallucination. ### Performance Comparison - On the Towel Folding task, the new model achieved 98.7% success, significantly outperforming baselines like POPE and PIO-0 - Baseline models struggled with tasks requiring context, often getting stuck in loops. ### Key Results - Asynchronous execution allows the model to maintain a smooth flow, predicting the next move based on the previous one, leading to a throughput of 1.0 pieces per minute vs. 0.5 for old methods - The model successfully generated smooth trajectories for complex actions like towel folding and Lego disassembly. ![Screenshot at 00:00: The title card and branding for the 'Really Easy AI' podcast featuring the 'Become a Member Today!' call to action.](https://ss.rapidrecap.app/screens/hmLkQ6Ob6Jo/00-00-00.jpg) ![Screenshot at 01:27: Visual representation of the concept being discussed, likely illustrating the challenge of real-time execution in robotics.](https://ss.rapidrecap.app/screens/hmLkQ6Ob6Jo/00-01-27.jpg) ![Screenshot at 02:24: A conceptual diagram showing the difference between noisy, discrete movements versus smooth, continuous motion.](https://ss.rapidrecap.app/screens/hmLkQ6Ob6Jo/00-02-24.jpg) ![Screenshot at 04:16: A slide or graphic mentioning the specific benchmark tasks used for testing, including Lego Disassembly and Towel Folding.](https://ss.rapidrecap.app/screens/hmLkQ6Ob6Jo/00-04-16.jpg) ![Screenshot at 09:06: A numerical display or chart indicating the 98.7% success rate achieved by the new model on tested benchmarks.](https://ss.rapidrecap.app/screens/hmLkQ6Ob6Jo/00-09-06.jpg)