UBTech brings customer production lines 1:1 to WRC, unlocking a real landing path for embodied AI

At this year's World Robot Conference (WRC), UBTech did not bring just one star robot — it moved real customer production lines onto its booth at 1:1 scale. The quickest way to judge whether an embodied-intelligence company has real capability is whether it dares to replicate a customer's actual line: promo videos can be edited and demos can be rehearsed, but a production line will not perform on cue.

On the floor, industrial humanoid Cruzr Y1 kept de-stacking and palletizing, moving cartons and totes, while Cruzr S2 loaded and unloaded automotive machined and sheet-metal parts with sub-millimeter positioning and picked cosmetics boxes of various sizes from mixed totes into a Dematic auto-seeding wall for sorting, at an average pace approaching 1,100 picks per hour. Everything — recognition, movement, grasping, placement — ran with no human intervention and no preset programs, and failed grasps were recognized and retried autonomously.

Jiao Jichao, UBTech's vice president and dean of its Embodied Intelligence and Humanoid Robot Research Institute, told QbitAI that customers evaluate robot vendors first on hardware stability, scene understanding, and whether a solution can actually run — not on benchmark scores. What customers pay for is the real scenario. The WRC setups have already passed the pilot-training stage and entered small-batch delivery, scaling up according to results in customer environments.

UBTech put all three product lines on stage at once: on the industrial side, nearly ten Cruzr S2 and Cruzr Y1 units worked continuously from opening to closing; on the commercial side, the new Walker C1 service humanoid targets reception, entertainment performance, R&D, and teaching assistance; and for home consumers, the ultra-bionic U1, powered by the Resonance-LM emotion model, focuses on emotional support and companionship.

The details show this was no choreographed show: automotive machining tasks require dual robots to carry parts over one meter long while handling incoming-material deviations, calibration drift, and on-site disturbance, landing final positioning accuracy below 1mm; sheet-metal loading runs the VLA model on the edge with the same accuracy; and e-commerce sorting demands sustained grasping of boxes in different sizes, colors, and stacking states.

UBTech is drawing a clear line between "working robots" and "dancing robots." Dancing robots favor lightweight bodies around 1.2 meters tall, planetary reducers, and chips like the Rockchip RK3588 that merely execute pre-scripted trajectories; working robots typically exceed 1.7 meters, use harmonic reducers to sustain loads and control contact torque, and need high-compute silicon such as NVIDIA Thor to run embodied and world models locally for real-time reasoning.

Supporting the continuous operation is a three-layer "embodied brain": the Thinker foundation model for understanding (nine first places in authoritative benchmarks for sub-10B embodied brain models, with the largest base model reaching 100B), the Thinker-WM world model for predicting action outcomes (built with Diffusion Transformer and Flow Matching and topping the LIBERO benchmark), and Thinker-VLA for turning task understanding into continuous control signals.

UBTech has pushed real-time capabilities to the edge: after optimization, Thinker-VLA inference efficiency rose 176%, storage usage fell 60%, and GPU memory needs dropped from 64GB to 32GB, with algorithms migrated from x86 to low-power ARM platforms so the whole stack can run on a single embedded board. The cloud keeps slow inference, multi-robot scheduling, and model management.

To adapt one "brain" across industrial, commercial, and home robots, UBTech uses a "1+N" strategy — one shared technology base and N scenario-defined products. The Walker C1 was built on the existing industrial robot base, reusing software and self-developed actuators, and took about four to five months from product definition to prototype versus six to nine months for a from-scratch build. Next to watch: the company's plan to fuse the three models into a unified end-to-end architecture, and whether small-batch deliveries can scale to hundreds or thousands of units.