Embodied AI — putting machine learning models inside physical robots — is usually framed as a model problem.
A Chinese robot called TA2 swept three industrial-skills events at this summer's WAIC competition in Shanghai and is now running shifts in a JD-owned 7Fresh supermarket in Taiyuan, one of the first open-retail deployments of "embodied AI," the term for AI that lives inside physical robots. Its maker argues the sweep proves the robot body, not the model, is the field's real bottleneck.
According to a leiphone.com analysis piece, the TA2 from Dexteleop (灵御智能) won all three categories at the inaugural "申智杯" embodied-AI contest during WAIC in July 2026: 213 picks in 15 minutes in dynamic sorting (about 4 seconds per pick), 121 assemblies in the parts event, and 29 box moves in materials handling. Remote-control entries were scored at half weight, and the article says TA2's converted sorting score still led the runner-up by 200-plus points. The numbers come from the company's narrative, not an independent scorecard audit.
The same piece anchors a commercial claim: TA2 robots are in routine operation at a Taiyuan 7Fresh, a setting the article calls rare for embodied-AI after five-plus rounds of pilot. Dexteleop's product page describes a unified control interface for teleoperation and model inference, with H.265 multi-sensor capture held under 10GB per hour.
The independent test is missing. No outside robotics researcher, competing body vendor, or second 7Fresh reference appears in the public record. The "body is the bottleneck" thesis is the vendor's to defend.