Xiaomi Tech

🎉 Congratulations to Daniel Povey, Xiaomi Distinguished Scientist, on being elected a 2026 ISCA Fellow. 🏆 ISCA Fellowship recognizes outstanding and sustained contributions to speech communication science and technology. Daniel's pioneering work on the open-source speech toolkit Kaldi has made foundational contributions to modern speech AI. At Xiaomi, he continues advancing the next generation of open speech technologies through projects and technologies including k2, Lhotse, Icefall, Sherpa, OmniVoice, and Zipformer. 👏 Congratulations on this well-deserved recognition! #XiaomiAI #SpeechAI #OpenSource #AIResearch #ISCAFellow

Scaling matters only if it transfers to real robots. We evaluate Xiaomi-Robotics-1 out-of-the-box on real-world mobile manipulation tasks -- including shoe organization, bag packing, table organization, and sofa tidying -- in unseen environments. Results showcase that scaling both pre-training data and model size consistently leads to better out-of-the-box performance in these real-robot evaluations. Furthermore, Xiaomi-Robotics-1 serves as a strong robot foundation model capable of adapting to novel tasks with minimal downstream data. Code and model checkpoints will be released soon.

Key findings: • Scaling data and model size consistently improve model performance during pre-training. • Scaling of pre-training transfers to strong out-of-the-box performance in unseen environments + efficient adaptation to novel challenging tasks. • #1 on RoboCasa365: 57.4% average success rate (vs. previous best 46.6%), with particularly strong generalization on the Composite-Unseen split. • #1 on RoboDojo: 20.07 average score & 13.93% success rate (vs. previous 13.07 & 8.80%). • 74.5% success rate on RoboCasa and 59.1% success rate on VLABench, outperforming RLDX-1, Cosmos Policy, GR00T N1.6, Pi-0.5, and Pi-0-FAST.

Can robot foundation models scale? Xiaomi-Robotics-1 explores this question with over 100,000 hours of real-world manipulation data. Pre-trained on large-scale real-world trajectories and post-trained with cross-embodiment robot data, Xiaomi-Robotics-1 demonstrates consistent scaling across data and model size, strong generalization in unseen environments, and efficient adaptation to new tasks. 🔗 robotics.xiaomi.com/xiaomi-robotic… #Robotics #EmbodiedAI #FoundationModels #RobotLearning #XiaomiAI #XiaomiRobotics

@reizzlla Thanks! We hope it helps push embodied AI forward. 🤖

@XiaomiTech_ This is massive for robotics 🤖

Introducing Xiaomi-Robotics-U0. A 38B unified embodied generation model for robotics. It brings scene generation, embodied transfer, robot interaction video generation, and image editing into a single foundation model—enabling scalable, controllable embodied training data generation. Built to generate scalable, controllable embodied training data for robot learning 🔗 robotics.xiaomi.com/xiaomi-robotic…

@SaniAiTech Exactly. Our goal is to make scalable synthetic data generation more accessible for robotics training and evaluation. 🤝

@XiaomiTech_ A unified foundation model like this could significantly accelerate robotics training and simulation

On the WorldArena leaderboard, Xiaomi-Robotics-U0 was evaluated under the anonymous submission "UNIS", ranking #1 among 126 models (as of July 15, 2026). We also evaluated Xiaomi-Robotics-U0 on real robot manipulation tasks, including earphone storage, towel folding, and box packing. Policies trained on data augmented with Xiaomi-Robotics-U0 remain more robust under unfamiliar backgrounds, challenging lighting, and other out-of-distribution conditions, while showing stronger self-correction during execution. We hope Xiaomi-Robotics-U0 serves not only Xiaomi Robotics, but the broader embodied AI community. #OpenSource #EmbodiedAI #Robotics #WorldModel #VLA

Key results: • 82.9× faster generation with FlashAR+ • #1 on the WorldArena benchmark (126 models) • +26.3% task completion in unseen environments • Outperforms GPT-Image-2.0 on multi-view structural consistency • Structured 5D controllable generation with geometric consistency across views • Supports multiple robot embodiments for scalable synthetic data generation

Our robot intern just reached another milestone. Earlier this year, Xiaomi's humanoid robot began its internship at Xiaomi EV Factory. After one quarter, its dual-side task success rate at the self-piercing nut workstation reached 98%—just 1 percentage point shy of the success rate achieved by experienced human operators. Now it's taking on two new factory roles: • Center console side-cover sorting • Returnable box folding Both have already achieved over 90% task success. Every new task is another step toward real-world deployment. #HumanoidRobots #Manufacturing #EmbodiedAI #Robotics #Xiaomi

"Find the photo of the person in red running by the beach." Current multimodal retrieval models see "person" but miss the red, the running, the beach. We call this Grain Blindness. We propose ELVA: upgrading contrastive learning with ranking-driven RL. On our new multi-grain benchmark MRBench, it improves by +13.1%. This work has been accepted to ECCV 2026. 📄 arxiv.org/pdf/2606.20280

Introducing Sky Nomad Xiaomi Sky Nomad reimagines the SUV as living space in motion, built for a generation that travels, works, and lives on the move. Officially unveiled in Mainland China.

Today we're announcing Xiaomi EV's second product series: Xiaomi SkyNomad. The Xiaomi SU7 and Xiaomi YU7 are positioned as "the driver's car". Xiaomi SkyNomad is positioned as "an intelligent, reconfigurable, large-space SUV". Different product lines, two answers to different user needs, delivered over more than five years of building cars.

How do you simultaneously overcome identity drift, temporal flickering, and lost details in degraded face videos? 🐯TIGER accepted at ECCV 2026 for identity-consistent face video restoration from severely degraded inputs. We fuse ArcFace identity embeddings, 3D cross-source parameter fusion, and a single-step rectified flow into one unified framework. SOTA on 8/9 benchmarks: LPIPS 0.07 (↓41%), FVD 40.4 (↓61%), IDS 0.95. 👇Visit our project page to see high-fidelity facial reconstruction. Project page + demos: yzhoulv.github.io/Tiger/ #ComputerVision #VideoRestoration #ECCV2026 #XiaomiAI

From restoring old photos to making cars drive safer — these 13 papers represent the full pipeline: perception → understanding → decision → action. More details coming soon

🚗 Autonomous Driving: 🔹 CausalDrive — real-time causal world model, 12 FPS 🔹 DriveVA — zero-shot cross-domain driving from video generation priors 🔹 MindDrive — online RL for VLA via language-space trial-and-error 🔹 SWAM — spatial-perceiving navigation, zero-shot generalization 🔹 BeyondDrive — learning safety from hard negatives 🔹 DriveFine — masked diffusion VLA with block MoE refinement

🏆 12 Xiaomi papers accepted at ECCV 2026!3 directions — vision understanding & generation, GUI agents & multimodal applications, autonomous driving 🧵👇#ECCV2026 #ComputerVision #XiaomiAI