What happened during Lexiang Technology (also known as JoyIn Technology)’s September 16, 2026 public Aether embodied-model stress test in Shanghai’s Minhang district, including the model’s approximate size, training data, technical approach, cross-embodiment capabilities, the rol

On September 16 in Shanghai’s Minhang district, Lexiang/JoyIn publicly streamed an outdoor barbecue-service test lasting more than an hour, presenting it as a real-world stress test of its Aether embodied model rather than a polished, pre-scripted demo. The evidence supports that Aether controlled robots through a sustained service workflow; it does not yet independently establish the company’s broader performance or generalization claims. 1215

• Model and training claim: JoyIn describes Aether as an approximately 4-billion-parameter embodied foundation model trained on about 200 hours of human video and “zero” real-robot training data. Its stated approach is perception/action-oriented physical interaction rather than a text-first system. These are company claims, not independently audited results. 715

• Cross-embodiment claim: The company calls Aether a cross-embodiment general model—one policy intended to work across differing robot bodies rather than requiring a separately trained model for each machine. Before the barbecue event, it drew attention with a claimed 10-minute, uncut, single-take video showing coordinated work across different embodiments. 1013

• What the robots did: Reporting describes a grill robot remaining at the barbecue station and cooking skewers, while service robots took orders, carried food, served customers, and interacted with them. The coordination highlight was a handoff: the grill robot passed finished food to a service robot, which brought it to a table. One service robot reportedly interrupted its current activity to fetch water when asked; others displayed more direct or more procedure-bound behavior. 12

• Why it was harder than a studio demo: The test was outdoors and open to real-time customer requests, interruption, changing objects and people, uneven environmental conditions, and a multi-step task chain—ordering, cooking, handoff, delivery, and interaction. Lexiang said it involved no rehearsal, preset script, or teleoperation; those autonomy conditions should be treated as assertions unless independently verified from full telemetry and operational logs. 101213

• What it establishes: A public, extended livestream is meaningful evidence that the demonstrated system could sustain some multi-robot food-service behavior in a less controlled setting than a short studio clip. It also makes failures, stalls, and recovery behavior more observable than edited footage. 1215

• What it does not establish: There is insufficient evidence to conclude that Aether is broadly robust, truly zero-shot across hardware, reliably autonomous without human intervention, or superior to competing embodied models. A reproducible technical report would need model architecture and interfaces, data provenance, hardware/control-stack details, intervention logs, failure rates, task definitions, held-out tests, and independent benchmark comparisons. None is supplied in the available reporting.