First combined continuous radar and camera run

On October 12, 2026, I ran firmware 0.3.2-rc1, build 87589b2c32638ae119334c72, with Radar Studio 0.9.0-preview.5 (hosted Site version 14) and a Logitech C920e. This was a brief physical acquisition and manual Stop check.

• Radar: 283 complete consecutive frames at 5 Hz, spanning 56.4 seconds, with no reported missed frames, capture faults, RF faults/errors or deadline misses. Matching-session keepalive acknowledgements were observed; manual Stop was acknowledged and followed by fault-free idle.
• Camera and recording: three fully decodable 1920 × 1080 video segments in a 68.27-second local session, alongside separate original control/data UART lanes. File hashes, byte lengths and lane totals matched. Segment gaps were recorded rather than treated as continuous exposure.
• Label review: the observed sit/stand/chair-roll sequence has approximate video-local timing only. Rolling a chair is not walking. These observations are not human-reviewed, synchronized training labels.
• Actual camera-model check: MediaPipe Pose Landmarker Lite processed a 29-second segment and returned unknown/unreviewed throughout because the view cropped the lower legs and feet. The importer correctly refused the unaligned review. No timing anchors or reviewed labels were invented.
• Training outcome: the existing single-detection baseline admitted 0 of 283 frames: 227 first failed truncated-detection coverage and 56 had multiple reflection peaks. Peaks are not people counts. The measurement exclusions were retained.

This owner-run test was reviewed by the development agent; it is not independent laboratory validation. It produced no trained radar model, held-out accuracy result or admitted dataset samples. Private recordings and video remain unpublished. Read the public aggregate run report.

Current software: continuous capture, local camera review and training preparation

Radar Studio 0.9.0-preview.5 supports manual timed or run-until-stop capture with the matching 0.3.2-rc1 firmware. Continuous operation uses an exact-session keepalive lease. Connection, configuration and pairing do not start RF; Start remains a separate manual action after exact-build/self-test, measured-mounting and no-RF port-pairing checks. The earlier 1–5 second mode remains available.

Local-folder recording preserves separate original control/data UART streams and optional camera segments. Any browser-supported camera can be selected; the C920e is the device tested in this run. Camera recording is optional, and recorded video does not establish automatic radar/camera synchronization. Captures are not automatically uploaded or admitted to the dataset.

The local replay and review workspace supports named sessions, annotations, replay and derived exports while keeping original observations and human decisions distinct. Google MediaPipe Pose Landmarker Lite supplies pose landmarks; separate experimental project rules propose activity intervals for human review. Poor framing, uncertain observations and missing alignment must remain unknown or unreviewed.

The separate host training tools 0.1.0-preview.3 provide optional Python validation, reviewed camera-label import, experimental baseline training and frozen-model evaluation. Import requires file integrity, matching firmware/session identity, human-reviewed labels and measured timing anchors, while retaining radar measurement exclusions. This is host software, not firmware to flash. Original-radar training is separate from compatible TI browser training; Python is not required for the viewer. No useful pretrained radar model or demonstrated activity accuracy is included.

Still ahead: deliberate keepalive-loss/disconnect stop and long-duration hardware endurance; timestamp rollover, phase/power calibration and motion-sign checks; full-body camera framing and measured radar/camera alignment; and eligible independent sessions and participant groups for training and held-out evaluation. Recording, review and export guide.

Earlier milestone: AOP polarity correction and device feedback

Firmware 0.3.1-rc1 corrects the documented fixed AOP receive-antenna polarity pattern. This is a hardware-convention correction, not measured phase or coordinate calibration. The matching Radar Studio 0.8.0-preview.2 adds separate control/data device terminals, clearer timeout recovery and official prerequisite links.

The earlier 0.3.0 hardware run accepted 25 frames and returned to idle, but XYZ estimates were rejected because the fixed antenna polarity had been omitted. At that release, the correction had passed software and compiled-target checks; physical coordinate acceptance remained pending. The October 12 continuous-run result above does not establish coordinate calibration. Candidate peaks and IDs still do not establish people counts, identities, posture or accuracy.

The viewer now offers a 20 m display with 2/5/10/20 m choices and close zoom. This changes only the viewport: the unchanged room64 sensing search covers approximately 0.8–5.9 m, and physical human-detection range is unvalidated.

Editable 1219 mm height and 0° tilt are starting values only. Antenna offset stays blank and measurement confirmation starts unchecked. Verify actual geometry before Configure. The viewer sends explicit runtime configuration after status and confirmation gates; a separate CFG upload is not required. At that release, Start was a separate, gated 1–5 second action after no-RF pairing. Browser layout and controls were checked without connecting USB or starting RF. Current capture modes and physical-run evidence are described above.

Previous release: dual-UART acceptance candidate

Radar Studio 0.8.0-preview.1 was paired with Original Radar Sensing 0.3.0-rc1. Enhanced control runs at 115200 baud; Standard data runs at 921600. Independent JSON streams, fresh cross-port pairing, exact-build/self-test checks and confirmed measured mounting gate a separate manual Start. Connection, configuration and pairing never start RF.

Acceptance plan at that release: first verify status/ACK and fresh cross-port pairing without RF, then stop for review before a separately authorized 1–5 second RF test. Physical dual-port routing, USB and RF sensing were still pending at that milestone; later acquisition evidence is summarized above. Calibration remains unverified.

The viewer adds local candidate association, sensor-coordinate 3D rendering, measured trails and experimental motion cues. The firmware supplies at most three exploratory peaks per frame. Candidate IDs and cues are not validated people counts, stable identities, postures or calibrated confidence.

At that release, the hosted edition displayed 0.8.0-preview.1. Its two-port controls, mounting fields, locked acceptance gate and 3D canvas render/reset were checked in a browser. Source, synthetic and offline checks passed; those checks do not replace hardware acceptance. Raw streams and exports remain client-local, with no contribution-upload service enabled.

Run Radar Studio in your browser

The hosted Radar Studio app is now public. Open it in compatible desktop Chrome or Chromebook Chrome with Web Serial support. USB access requires you to select and grant access to the correct serial ports. Use the workflow and profile that match the firmware already on your board; opening or updating the app does not flash firmware.

Recording and analysis run client-side. Captures stay local unless you explicitly export or share them; the app does not automatically upload recordings into the dataset. The original-firmware path still requires the exact supported build and status/self-test checks before a separate manual Start in the selected supported capture mode. No automatic RF start is performed.

Hosted-app check: the interface, 12-profile catalogue, original-firmware controls, offline route and guide loaded in a browser, and a synthetic demo ran and stopped. That does not validate every demo, physical USB operation, RF sensing or real-world accuracy.

Open offline analysis · Read the browser guide

Choose the correct capture path

Existing TI-compatible viewer workflow

Use Radar Studio with the exact compatible TI firmware and protocol profile already on the board. The original prototype had user-reported live tracking and vitals; newer viewer releases and additional profiles still require browser and hardware acceptance testing. Received UART capture preserves what the firmware transmits; it is not raw ADC capture.

Official dependencies and matching TI package paths

Original sensing workflow and earlier collector

Original Radar Sensing uses its own JSON Lines protocol. The earlier Python collector describes the historical single-port workflow; it is not a matching controller for dual-UART v2 firmware. Use the current matching viewer for the new candidate. Its configuration and observations are separate from TI CLI/TLV data and must not be silently fed into a TI-trained model. It is optional experimental bring-up work, not a prerequisite for the existing viewer.

Original collector setup and operation guide

What the evidence supports today

• Released software and offline tests establish specific code behavior. They do not establish real-data activity accuracy, browser rendering or hardware throughput.
• Training uses explicitly selected captures, with independent held-out sessions for evaluation. Uncertain observations can remain unknown; human review makes the final annotation decision.
• The original sensing firmware has no completed activity classifier or demonstrated tracking accuracy. Earlier 25-frame bring-up checks have now been followed by the October 12 283-frame continuous acquisition and manual Stop check. Deliberate heartbeat-loss/disconnect testing, long-duration endurance, calibrated measurements, synchronized reviewed labels and independent held-out activity evaluation remain ahead.
• This work is experimental. Vitals from compatible TI firmware are not medical measurements, and these tools must not make medical or safety decisions.

Where the dataset could lead: a smarter home

I want labeled sitting, standing, walking and transition recordings to help train and evaluate future room-occupancy and activity models. A few practical possibilities:

These are proposed applications. The project does not currently actuate lights, TVs or other home devices. The public dataset still has zero released samples, and there is no validated people tracker or demonstrated activity-classification accuracy.

A useful automation would need confidence thresholds, debouncing across repeated observations, a time delay and a manual override. Brief signal gaps should not switch off someone’s lights. Reliable handling of stationary people, room-specific testing and clear household consent would come before deployment.

Contribute recordings or help review

Contact dataset@swarmcord.com first with a short, non-sensitive summary of your proposed recording or volunteer contribution. Please arrange an appropriate private transfer and participant permissions before sending captures or video. Do not put personal or sensitive recordings in public GitHub issues.

Useful submissions identify the firmware and configuration, export schema, session IDs, file hashes, human-observed labels and event boundaries, quality flags, mounting and timing uncertainty, and consent/reuse scope. Large files need an agreed transfer location.

I review submissions and manually add accepted recordings. Nothing is accepted or published automatically. Contributors can request opt-in name or handle credit; anonymous participation is also available.

Source availability and reuse

The original software is source-available under the project’s no-resale license, not OSI-approved open source. TI and other third-party components retain their own terms, including applicable device restrictions. Use complete release packages with their required notices.

Dataset terms are separate and remain pending. The intended future policy includes credited research, teaching and commercial AI training, but that intent does not grant rights to unreleased recordings. Approved releases will carry their final terms.

Software license · Firmware licensing and notices · Back to all projects