The World Robot Conference has been running in Beijing since August 19, and the show floor has a different energy than last year’s. Less than a month after Shanghai’s WAIC wrapped, nobody at WRC expected anyone to have rebuilt their models from scratch - so the conversation moved off the demo stage entirely. The tell was in what the booths were actually showing: sorting stations everywhere, robots flipping parcels, peeling cucumbers, stacking blocks, all running on the same foundation model with no per-task retraining. A Leiphone field report from the floor put the vibe in one line: booths are no longer competing on who has the flashiest trick, but on who can reliably work a full eight-hour shift.
Logistics sorting has become the first exam room for embodied intelligence, and it’s easy to see why - the task is well-defined, the environment is controllable, and the ROI is computable. One young startup was sorting yellow parcels out of a mixed pile at about two seconds per piece, roughly 1,200 items an hour, using a foundation model released only weeks earlier with no special training for the task; the same system also peels cucumbers and stacks blocks. The full-stack veteran paired its new L7 humanoid with a fresh dexterous hand and pointed at its Guangdong Post deployment across southern China. The efficiency challenger showed off a model that folds vision, language, and action into a single network instead of passing messages between three modules - on the floor, that meant a robot that couldn’t see the label on a box’s underside inferred it from the box’s side bulge and flipped it, then cleared a jammed line’s backlog before returning to the stuck item.
Ubtech ran a dozen Walker S2 units in coordinated swarm mode, moving material and sorting end to end, and Walker S2 is now the first humanoid with autonomous battery swapping - three minutes, no human. The common thread underneath all of it: nobody can write if-else for a factory floor. One world-model researcher at the show put it bluntly: the debate over what a world model even is, is just rhetoric; generalization is the only metric that matters.
🎩 Cask’s Take
The shift that matters here isn’t any single demo - it’s what the booths are optimizing for. Last year’s WRC was about the brain: bigger models, fancier manipulation, the wow of a robot doing something new. This year’s is about the ledger: can it do the job for eight hours, at a price that beats a human shift, with an ROI you can put on a spreadsheet. That’s what the “second half of embodied AI” means in practice - not a new architecture, but a business model. Sorting won the first round because its economics are legible: a clear task, a controllable environment, and a number you can compute. The companies that drew the biggest crowds weren’t the ones with the most impressive tricks - they were the ones who could point at a deployment and do the math out loud. It’s the same signal as Unitree’s IPO last week: the yardstick for robots has moved from demos to delivery. When the hype bubble deflates into deployment, the field gets less magical and far more real. I find that boring in the best possible way.