NVIDIA/xr-ai v0.2.0-beta
NVIDIA/xr-ai
Captured source
source ↗published Sep 9, 2026seen 12hcaptured 12hhttp 200method plain
XR AI v0.2.0 Beta (Withdrawn)
Repository: NVIDIA/xr-ai
Tag: v0.2.0-beta
Published: 2026-09-09T00:25:09Z
Prerelease: yes
Release notes: > [!WARNING] > This release is withdrawn. The immutable v0.2.0-beta tag was created from main instead of the intended release branch. Do not use it. Use v0.2.0-beta.1 instead. > > The notes below describe the intended beta and do not match the withdrawn tag contents.
This beta introduces a revised agent architecture and two new end-to-end samples:
- Adds typed AgentRuntime and VoiceAgent components with Relay-managed native tools. In-tree MCP servers and public Pipecat pipeline assembly APIs are retired.
- Adds tea-making guidance with evidence-gated workflow progression, background observation, document retrieval, and durable artifacts.
- Adds lab instrument monitoring with marker-associated display readings, change alerts, durable output, and live event viewing.
- Separates image selection from inference, supporting current and historical frames, recorded windows, multi-image requests, and timestamped video-frame queries.
- Simplifies model setup by letting agent samples share the same prestarted speech, language, vision, and embedding services, with configurations for local models and NVIDIA NIMs.
- Improves participant-scoped voice output, interruption handling, service readiness, model service reuse, and cross-platform network telemetry.
- Publishes versioned documentation with generated SDK, command-line, and configuration references.
This release contains breaking package, API, and configuration changes, including the DeviceIOHub rename. Consult the release migration guide when updating from the previous beta.
Full Changelog: https://github.com/NVIDIA/xr-ai/compare/v0.1.0-beta...v0.2.0-beta
Known issues
- The web client retains the previous agent response after disconnecting.
- The iOS client may remain reconnecting for about a minute after server shutdown and then report a track-removal error.
- XR Render discards conversation context older than ten minutes, even during an uninterrupted session.
- XR Render deployments using recorded video must explicitly enable video_history_enabled; omitting it disables historical queries.
- With video history disabled, historical questions may incorrectly query the current frame instead of reporting that history is unavailable.
- The XR Render live-perception evaluator may accept a response that combines an inability statement with a fabricated physical claim. It may query the scene for a user query instead of the camera in certain cases.