Stats
Actions
Tags
From aria-ark
Use when working with Aria Machine Perception Services (MPS) — cloud-based processing of Aria VRS recordings for SLAM trajectories, hand tracking, semi-dense point clouds, and online calibration. Covers the aria_mps CLI, single and multi-sequence processing, output formats, data lifecycle, and loading MPS results with ProjectAriaTools. Use whenever the user mentions MPS, aria_mps, SLAM processing, cloud hand tracking, trajectory generation, or submitting VRS files for processing.
How this skill is triggered — by the user, by Claude, or both
Slash command
/aria-ark:mpsThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
MPS is a cloud-based post-processing service for Aria recordings. You submit VRS files via the `aria_mps` CLI, MPS runs proprietary Spatial AI algorithms, and you download the derived outputs (trajectories, point clouds, hand tracking). The data is only used to serve requests — not accessible to Meta researchers.
MPS is a cloud-based post-processing service for Aria recordings. You submit VRS files via the aria_mps CLI, MPS runs proprietary Spatial AI algorithms, and you download the derived outputs (trajectories, point clouds, hand tracking). The data is only used to serve requests — not accessible to Meta researchers.
This skill teaches concepts and capabilities. For details, use the authoritative sources:
aria_mps --help, aria_mps single --help, aria_mps multi --helphelp(projectaria_tools.core.mps) or MpsPyBind.h in the cloned repo. See the projectaria-tools skill.pip install projectaria-mps (automatically includes projectaria-tools and projectaria-vrs-health-check)aria_mps$HOME/.projectaria/mps.ini — concurrency, chunk sizes, retry logic.MPS provides three services. Services are independent — you can request any combination. VRS health check runs automatically before processing.
Check the versioning docs for current service versions and the benchmarks page for accuracy metrics.
aria_mps single): Process recordings independently. When pointed at a directory, recursively discovers all VRS files and processes concurrently (default 25, configurable up to 100).aria_mps multi): Process multiple recordings together for Multi-SLAM in a shared coordinate frame.--features: Select specific services (e.g. SLAM, HAND_TRACKING). Default: all applicable.--force: Bypass deduplication cache and reprocess.--no-ui: Headless mode for scripted/automated pipelines.The CLI executes stages sequentially: status check → health check → encryption → upload (resumable within 24h) → server processing → download. Quitting after upload lets processing continue in the background — results download automatically when you rerun.
Results are saved alongside the input VRS file in a sibling folder. The structure includes separate directories for each service (slam/, hand_tracking/) plus a VRS health check report. Each service folder contains data files (CSV/JSON) and a summary.json with quality metrics.
See the data formats docs for exact CSV column definitions, coordinate conventions, and timestamp semantics.
tracking_timestamp_us (device clock, monotonic, stable) is for time calculations. utc_timestamp_ns (wall clock) is NOT monotonic — avoid for durations.graph_uid: Identifies the world coordinate frame. Data sharing the same graph_uid is spatially consistent.quality_score: Per-pose confidence (0–1). Filter low-quality poses for downstream use.uid for per-frame 2D-3D correspondences.--force to bypass.On-device eye gaze and VIO are embedded in the VRS during recording (accessible via PAT Tutorial_4 and Tutorial_5). Cloud MPS produces substantially better results using offline algorithms — higher accuracy for both SLAM and hand tracking. See the benchmarks page for quantitative comparisons.
When any MPS command fails, returns unexpected output, or produces low-quality results, you MUST consult this section and the official troubleshooting page BEFORE attempting your own diagnosis. This is not optional. MPS failures have MPS-specific root causes (health check rejections, insufficient recording length, coordinate frame mismatches) that generic debugging will misdiagnose.
Anti-patterns:
--force wastes upload time and quota.Official troubleshooting page: https://facebookresearch.github.io/projectaria_tools/gen2/ark/support/mps — covers symptoms not listed below.
summary.json for per-service status (SUCCESS/WARNING/ERROR) and quality metricsnpx claudepluginhub facebookresearch/projectaria-plugins --plugin aria-arkProvides behavioral guidelines to reduce common LLM coding mistakes, focusing on simplicity, surgical changes, assumption surfacing, and verifiable success criteria.
Searches, retrieves, and installs Agent Skills from prompts.chat registry using MCP tools like search_skills and get_skill. Activates for finding skills, browsing catalogs, or extending Claude.
Creates, edits, and optimizes skills for Claude Code, including drafting, evaluating with test prompts, iterating on performance, and improving skill descriptions for better triggering accuracy.