Less than a month after WAIC in Shanghai, Beijing hosted WRC 2026. Nobody expected a model rewrite in a few weeks, so this field report reads the show against months of company performance to see where the ’embodied brain’ actually stands.
The clearest signal: professional visitors stopped admiring prototype demos and started judging whether the brain can work steadily and whether the economics close. Logistics sorting is the first commercial scene to be proven and most rehearsed, with clear tasks, controllable environments and calculable return on investment, but now crowded.
Sorting as the exam hall
ForceLink showed fruit cutting and a two-arm parcel sort at about 2 seconds per item, or 1,200 items an hour, using its DM0.5 generalisation model launched on 9 July, with no retraining across sorting, cucumber cutting and block stacking. Astribot, the ‘old hand’ of sorting, ran dexterous-hand sorting on a real conveyor with its new L7 body and XHAND 1 PRO hand, and pointed to work with Guangdong Post. Zhiyuan Robotics bet on efficiency with its WALL-B model (24 July) that fuses vision, language and action in one network, even flipping a parcel to find a hidden label.
From single robot to swarm
UBTECH staged more than ten Walker S2 humanoids collaborating through its group-brain network 2.0, doing material handling, sorting and transfer as a cluster; Walker S2 is the world’s first humanoid with self-swapping batteries, a 3-minute autonomous swap aimed at the endurance problem. Galbot pushed its ‘brain plus cerebellum’ story: AstraBrain-WBC 0.5, called the world’s first general ‘cerebellum’ GPT base model, with 0.39-millisecond inference latency for zero-shot basketball, boxing and dance, paired with AstraBrain WAM 0.5, a world-action model, inside the end-to-end AstraBrain running on the wheeled dual-arm Galbot G1 for long, messy home and retail tasks.
RoboScience showed its REX G1 wheeled humanoid (17 August) with millimetre-level control. XSquare’s Moz1 improved living-room tidying efficiency by more than 15 per cent in 30 days of full-stack iteration. DexForce ran one brain across phone-line inspection, supermarket scanning and ice-cream service, replicating a Lens Technology production cell on the show floor. Wujan Power ran a robotic coffee space with Korea’s HOLLYS and cited RMB 700 million in cumulative orders with global delivery started.
The infrastructure layer
Octopus Dynamics, founded after late 2025, raised about RMB 1 billion in three months and showed SYNWorld V0.5, a unified multimodal world model that visualises its own predictions. LingChu demonstrated wearable ego-data capture mapping a demonstration to a robot in about 10 seconds, with 100,000 hours of human operation data banked. Lumin’s MOS2 heavy-load wheel-arm robot feeds a NexCore learning loop. GigaAI’s synthetic-data GigaWorld reaches 90 per cent of VLA training with roughly 300 per cent gains on three generalisation axes. Qiongche’s AnyGrasp library posts a 99.8 per cent single-grasp success rate.
The National and Local Co-built Humanoid Robotics Innovation Centre skipped the demo and shipped open-source bases and unified training and evaluation platforms, arguing the industry needs a ruler for generalisation now that embodied-model leaderboards exceed twenty and claims of ‘world first’ are cheap.
SongYan leaned into expressive faces and anime IP to dodge the uncanny valley, and reported three number-one finishes in JD’s June 18 humanoid category, a RMB 1 billion Series B, and education orders above half. The blunt read: the industry is moving from ‘PPT plus backflip’ to ‘scenario plus model plus can-it-profit’. Every firm is hunting product-market fit and a closed loop in the real physical world.
Editor’s note: This is an adapted translation of the original LeiPhone report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://www.leiphone.com/category/ai/rUNBvF7Iq6C3UtyR.html.