Xiaomi Robotics announced on August 18 that it won first place in two major international competitions within the same time window: the RoboChallenge Track at the CVPR 2026 Workshops GigaBrain Challenge, and the Whole Body Control (WBC) competition at ICRA 2026. This marks the first time a Chinese embodied AI company has simultaneously claimed championships in two very different dimensions — "single-model multi-task real-robot complex manipulation" and "whole-body coordinated control" — in competitions with noticeably raised difficulty curves and scoring standards compared with previous years.
The CVPR 2026 RoboChallenge Track tests single-model, multi-task real-robot manipulation: systems must reliably execute long-horizon instructions across task families on the same robot platform, covering assembly, household chores, and industrial-grade fine manipulation. Xiaomi's anonymously entered my16 model topped the leaderboard with a 40.89% success rate, becoming the only model in this edition to break the 40% line — the gap between second and third place is far smaller than the gap to the 40% line itself.
my16 is not a specialist trained for a single task. It uses a "VLM brain + world-model cerebellum" dual-system architecture: the VLM brain handles long-horizon task understanding (vision-language instruction parsing, task decomposition, step generation), while the world-model cerebellum handles fine manipulation (contact dynamics prediction, motion control, force-feedback closed loops). The two systems are coupled within a single end-to-end policy network, giving the robot both the ability to "understand human language" and to "physically touch things."
The ICRA 2026 WBC competition tests a completely different curve: whether robots can reliably perform whole-body coordinated control, including bimanual arms, torso posture adjustment, a mobile base, and whole-body dynamics under human-robot interaction. Xiaomi Robotics took first place with a 99.2 composite score and 94% overall success rate, the only team above 90% this year — leading second place by 10 percentage points. The engineering challenge lies in the word "whole-body": when two arms execute different tasks simultaneously, dynamic coupling amplifies per-joint control errors, and Xiaomi's whole-body coordination policy absorbs those errors inside the control loop.
What the 40% breakthrough means
40.89% may not look high, but RoboChallenge evaluates long-horizon tasks on real robot hardware, with assembly, household, and industrial scenarios that were not part of training. Second place likely sits in the 30–35% range; the 40% line means my16 has a clear advantage across the full chain of "understand the instruction → decompose steps → execute across tasks." The VLM-brain + world-model-cerebellum architecture essentially solves "task-level understanding" and "contact-level control" with two different model types, then fuses them at the policy level — an engineering inflection point repeatedly validated in embodied AI over the past year, and a high ground every vendor is competing for.
Why 94% is harder than 40%
The 94% overall success rate at ICRA WBC was achieved in the whole-body dimension — bimanual arms, torso posture, and mobile base working simultaneously — far harder than single-arm tabletop tasks. The 99.2 composite score is a weighted result across sub-dimensions including motion smoothness, energy efficiency, and error recovery, as judged by the evaluation committee. Being the only team above 90% implies second place likely landed in the low 80s. Viewed together, the "VLM brain + world-model cerebellum" architecture is not just a shortcut to a single title, but a general framework that achieves engineering-optimal performance in two distinct dimensions: long-horizon tasks and whole-body coordination.
An engineering inflection point for Chinese embodied AI
The timing of the dual championship is notable — CVPR 2026 and ICRA 2026 ran in the same window, with major Chinese embodied AI vendors competing anonymously, raising the level of competition to a new bar. Xiaomi winning both very different tracks suggests its embodied AI development has shifted from "single-point breakthroughs" to a "general platform." Combined with AtomBrain's nearly RMB 1 billion Series A (Aug 15) and JD's RoboBase 80+ five-year plan (Aug 13), the "academic results → engineering deployment → capital validation" loop for Chinese embodied intelligence is closing fast.
What it changes
The dual championship leaves the industry with three takeaways: first, the "VLM brain + world-model cerebellum" dual-system architecture has been validated on international tracks as a scalable engineering paradigm; second, the sole breakthroughs of 40% on long-horizon tasks and 94% on whole-body coordination settle the question of whether domestic embodied robots can handle complex tasks; third, Xiaomi has anchored its robotics R&D to a clear target of "factory production lines + the real world" — the last mile between academic results and commercial deployment. The next round of competition in Chinese embodied AI will shift from championship contests to scaled delivery.