On August 18, Xiaomi's robotics team announced that it had taken first place in two major international competitions held within the same time window: the CVPR 2026 Workshops GigaBrain Challenge's RoboChallenge Track and the ICRA 2026 Whole Body Control (WBC) competition. According to the post, this is the first time a Chinese embodied-AI company has won championships in two fundamentally different dimensions—'single-model multi-task real-robot complex manipulation' and 'whole-body coordinated control'—during the same period, in a year when both difficulty curves and scoring standards were notably raised.
CVPR 2026 RoboChallenge: 40.89% Success Rate, the Only Model Above 40%
The RoboChallenge Track tests single-model, multi-task complex manipulation on real robot hardware. Participating systems must reliably execute long-horizon instructions across task families on the same robot platform, spanning assembly, household chores, and industrial-grade fine manipulation. Xiaomi's anonymously entered my16 model topped the overall leaderboard with a 40.89% success rate, the only model in this edition to break the 40% threshold—the gap between second and third place is far smaller than the distance to the 40% line itself.
my16 is not a single-task specialist. It follows a dual-system architecture:
- VLM 'brain': long-horizon task understanding, including vision-language instruction parsing, task decomposition, and step generation
- World-model 'cerebellum': fine manipulation, including contact dynamics prediction, motion control, and force-feedback closed loops
ICRA 2026 WBC: 99.2 Points and 94% Overall Success Rate
WBC tests a completely different skill curve: whether a robot can reliably perform whole-body coordinated control, including dual-arm manipulation, torso posture adjustment, mobile base locomotion, and whole-body dynamics under human interaction. Xiaomi Robotics scored 99.2 points overall with a 94% overall success rate, ranking first and standing as the only team in this edition of WBC above 90%—a 10 percentage point lead over second place.
The engineering difficulty lies in the word 'whole-body': when dual arms execute different tasks simultaneously, dynamic coupling amplifies control errors at individual joints. Xiaomi's whole-body coordination strategy absorbs these errors within the control loop itself.
What the 40% Breakthrough Means
While 40.89% may look modest, the RoboChallenge Track evaluates long-horizon tasks on real robot hardware, with a test set including assembly, household, and industrial scenarios not seen in training. Second place likely falls in the 30%–35% range. Crossing 40% means my16 holds a clear advantage across the full chain of 'understanding instructions → decomposing steps → executing across tasks.' The VLM-brain + world-model-cerebellum architecture fundamentally solves 'task-level understanding' and 'contact-level control' with two specialized model types, then fuses them at the policy layer—an engineering inflection point repeatedly validated across the embodied-AI field over the past year, and high ground every vendor is racing to claim.
Why 94% Is Harder Than It Sounds
The 94% overall success rate was achieved in the whole-body coordination dimension—dual arms, torso, and mobile base working simultaneously—a far harder task profile than single-arm tabletop manipulation. The 99.2 composite score reflects the evaluation committee's weighting of multiple sub-dimensions, including motion smoothness, energy efficiency, and error recovery. Being the only team above 90% implies second place likely sits in the low 80s. Viewed together, the two titles suggest the dual-system architecture is not a single-event shortcut but a general framework achieving engineering optima across both 'long-horizon tasks' and 'whole-body coordination.'
An Engineering Inflection Point for Chinese Embodied AI
The timing is notable: CVPR 2026 and ICRA 2026 opened in the same window, with major Chinese embodied-AI vendors all competing anonymously, raising the intensity of competition to a new level. Winning both tracks signals that Xiaomi's embodied intelligence R&D has shifted from 'single-point breakthroughs' to a 'general platform.' Together with other recent milestones—AtomBrain's nearly 1 billion RMB Series A (Aug 15) and JD's RoboBase five-year plan exceeding 80 (Aug 13)—the 'academic results → engineering deployment → capital validation' loop for Chinese embodied intelligence is rapidly closing.
What Changes
Xiaomi's double championship leaves the industry with three takeaways:
1. The 'VLM brain + world-model cerebellum' dual-system architecture is validated as a scalable engineering paradigm on international tracks. 2. Being the only entrant to break 40% on long-horizon tasks and 94% on whole-body coordination removes doubt about whether domestic embodied robots can handle complex tasks. 3. Xiaomi is anchoring robotics R&D on the concrete goal of 'factory production lines + the real world'—the final stretch between academic results and commercial deployment. The next round of competition in Chinese embodied AI will shift from championship contests to large-scale delivery.
*Source: Xiaomi Robotics team announcement, August 18.*