Context
The fleet, six named robot models, each task-specific:
- XMAN-R1, humanoid greeter (front-desk welcome, natural-language Q&A, welcome-gift presentation) - W3, in-room delivery (amenity drop-off) - S100, high-payload luggage transport - C40, autonomous cleaning - T10 and T3, F&B delivery within the restaurant
The architectural inverse of Henn-na: multi-unit specialisation rather than one form factor doing everything.
What the AI actually does
Operating logic is explicitly general-purpose + special-purpose collaboration, humanoid handles unstructured guest interaction, special-purpose units handle high-volume repetitive tasks. Routes for the delivery robots are deterministic (mapped paths); the humanoid uses VLA for greeting interactions.
Whether humans remain in the loop is not publicly described in the launch coverage; treat as architectural inference, not verified differentiator.
Measurable outcomes
- No KPIs, no cost figures, no guest-satisfaction data, no supervision-protocol details, and no executive on-record quotes from either Shangri-La or Keenon were published in any of the seven canonical launch sources
- No 2026 follow-up coverage of operational performance has appeared as of April 2026
- Confirmed gap, not a research oversight
What to copy
The structural choices distinguishing Shangri-La/Keenon from Henn-na are task decomposition and vertical fine-tuning on top of a foundation model, not the humanoid form factor itself.
Adopt the pattern; treat the marketing as marketing until the first independent guest-experience and ROI data lands. Physical AI can work in hotels, but only when the operating logic is "free humans from mechanical work" rather than "replace humans".
What doesn't transfer
Vendor superlatives ("world's first general-purpose + special-purpose hotel collaboration model") do not transfer. The architecture is credible but the outcome is unverified.
Foundation-model + vertical-fine-tune separation is a 2025-era pattern that did not exist in 2015–2018; the floor-plan lock-in problem is also reduced because the robots are mobile fleet units, not building-fixed installations.
Open questions before buying
- What is the human-supervision protocol per robot model? It is not publicly described.
- Where do failures escalate? Who authorises a reset, a retraining episode, a withdrawal of a unit?
- What is the reported guest-satisfaction delta across the deployment window?
- What is the per-unit total cost of ownership including software updates over the depreciation window?
The vendor, Keenon Robotics
Keenon claims 15 years, 100,000+ deployed units, and 60+ countries (vendor claim, unverified independently).
KOM 2.0 is Keenon's self-developed Vision-Language-Action (VLA) foundation model, marketed as "the world's first VLA model for the service industry" (vendor claim). KEENON ProS is the vertical fine-tuning layer that gives robots hospitality-specific "vocational skills" on top of KOM 2.0.