Ant Lingbo CEO: robots can't even scale to corner stores yet — data pipeline is the real moat
APPSO · wechat · 2026-09-16
At the 2026 Bund Summit, APPSO sat down with Zhu Xing, CEO of Ant Group's embodied AI unit Lingbo, for a candid look at why robots still aren't taking over everyday work.
Reality check: neither capability nor cost is ready — even a corner store is hard to deploy at scale. Current deployable products rely mostly on imitation learning, which breaks on unseen anomalies. Deployments need constrained environments and short-horizon tasks; copying a solution to a new pharmacy or store remains expensive due to poor generalization. Their pharmacy picking system is already live in Guoda Pharmacy stores, navigating 80cm-wide aisles.
Why train a foundation model from scratch: Lingbo first fine-tuned general video models (e.g. Wan) into world models but hit a wall — low ceiling and destroyed priors. They now train LingBot-Video and the action model LingBot-VLA2.0 natively with real robot data. Robots need physically plausible, real-time prediction, not pretty visuals. The base model's value is better few-shot post-training: same results with less data.
The data recipe: Zhu believes in scaling laws but not naive data volume — he's seen 3-5x more data perform worse due to distribution imbalance. Lingbo moved to custom data collection with a pipeline pushing usable data from 15% to 90%+, using streaming delivery. Internet video suits pretraining; real-robot and sim data matter most in post-training. Ego capture is promising but hand-trajectory accuracy is still too low; he favors high-precision haptic gloves.
Other takes: the real data flywheel runs on anomaly data, not successes; everyday human data is "coarse grain" robots can't eat yet; LLM companies entering robotics won't automatically win — the moat is data understanding, like self-driving; humanoids needn't be human-shaped, form should follow task and cost; home robots will progress stepwise from companionship to assistive manipulation, with no timeline for full trust.
More from Embodied
- Drone caught on camera giving a dog a bath in China — ZeroStateReflex · 2026-09-16
- RoboDojo Puts GPT-6 Astra to the Test: Strong Semantics, Weak Physical Commonsense — ZeYanjie · 2026-09-16
- Muse Now Picks Its Own Outfit Every Morning, Scale CEO Approves — alexandr_wang · 2026-09-16
- Scoble predicts billions of evolving personas per person, ahead of Meta Connect avatar demo — Scobleizer · 2026-09-16
- Hacked computer-use for iPad: USB dongle bypasses iPadOS accessibility limits — alfcnz · 2026-09-16
- Apple's Reference Image hashes raw negatives in-sensor for verifiable photography — nptacek · 2026-09-16