A kiosk avatar lives in a noisy, public, shared, and often unattended environment. The right platform must fit fixed hardware, constrained networks, microphone and speaker realities, privacy, accessibility, session reset, remote monitoring, and a fallback that still completes the task. A polished office demo is only the beginning.
Evaluate one complete site, not one browser tab. Include boot, idle, peak footfall, background noise, privacy reset, network loss, remote update, and an operator recovering the device.
| Criterion | What to evaluate | |
|---|---|---|
| Hardware fit | OS, CPU/GPU, memory, display, camera, microphone array, speaker, thermal envelope, peripherals, and locked-down browser/runtime. | |
| Site network | Shared uplink, firewall, proxy, packet loss, offline behavior, asset caching, traffic per kiosk, and reconnect storms after outage. | |
| Public interaction | Wake word or touch start, barge-in, captions, language selection, volume, bystander speech, privacy, and clear session ending. | |
| Task integration | Directory, appointment, ticketing, payment, badge, accessibility device, printing, and safe rollback after a partial transaction. | |
| Fleet operations | Provisioning, health, remote logs, content and SDK update, secrets, certificate rotation, alerting, support, and physical reset. | |
Client rendering can reduce steady-state avatar traffic, while cloud video can reduce device rendering work. The site’s hardware and network decide which trade is better.
| Platform | Product boundary | Strongest fit | What to verify |
|---|---|---|---|
| Spatius | Client-rendered avatar with compact motion delivery | Kiosks with capable edge hardware, constrained uplinks, and a customer-owned agent | Validate GPU/CPU, asset caching, thermal stability, and full offline fallback |
| D-ID | Web-oriented streamed agents | Kiosks using a browser-based D-ID agent experience | Test codec, firewall, resolution, microphone publishing, session reset, and remote recovery |
| Tavus | Managed conversational video interface | High-touch kiosk experiences where managed video is central | Model sustained site egress, concurrency, camera/mic permissions, and vendor outage behavior |
| Anam | Managed conversational persona | Rapid web kiosk prototypes | Verify locked-down browser support, noisy-room behavior, privacy controls, and fleet observability |
A useful pSEO comparison should make the decision reproducible, not merely repeat vendor language.
This boundary prevents an avatar-runtime claim from being mistaken for a complete product outcome.
The deployer owns hardware, site network, microphones, speakers, ASR, LLM, TTS, domain tools, transaction safety, privacy notice, physical accessibility, session reset, secrets, fleet monitoring, updates, onsite support, and offline/fallback workflows.
Spatius supplies the motion service and client AvatarKit runtime. Its low-byte motion path can help at constrained sites, but Spatius does not manage kiosk hardware, speech capture, transactions, fleet health, physical privacy, or the complete network footprint.
The same platform can be an excellent layer for one team and the wrong amount of infrastructure for another.
Text and touch are better for noisy places, private data, precise selections, and users who do not want to speak in public. Voice-only can be better for eyes-busy or low-vision interaction when the screen or GPU is constrained. Offer simultaneous captions and a clear touch fallback.
A human is better for identity disputes, payment problems, accessibility assistance, safety incidents, complex exceptions, and repeated failure. Provide a visible call button or staffed escalation rather than trapping users in the kiosk.
Run a seventy-two-hour site test on production hardware, then simulate network, power, peripheral, and backend failure.
Reviewed Aug 3, 2026. Product modes, plan limits, pricing, and documentation can change. Recheck every source before purchase or publication. Sources establish platform capabilities; the selection framework is Spatius editorial analysis.
