v0.8.2

• Publish the operator-approved, installation-specific Astra Pro RGB head-camera
  baseline 20261110-v1: native 1280×720 intrinsics/distortion, metric
  base_from_camera, coordinate conventions, robot FK identity and provenance.
  A SHA-256-pinned current selection prevents silent parameter replacement.
• Bind the same bytes to the existing Hugging Face OpenArm supplied-gripper model
  in asset package v0.3.1. The immutable lock and camera scene entry reproduce
  head_calibrated without modifying original meshes, task cameras or model/data
  payloads. Retain physical-calibration limitations and the original raw evidence
  externally; this is not independent absolute-accuracy certification.
• Add dependency-free configuration integrity tests and include the baseline and
  asset lock in source distributions. No motor behavior, device selection,
  live-service deployment or automatic hardware activation changes.
• Fix a read-only collection-preview freshness race by sampling wall time after
  reading an atomic status snapshot. Preserve explicit-time checks and stale/future
  rejection; add a deterministic concurrent-publication regression test.

v0.8.1

• Add same-origin collection visualization at /#collection: fresh/stale
  recording status, stream activity and writer queues, plus read-only episode
  inspection, causal camera/tactile/joint playback and acceptance/quality labels.
  A best-effort status publisher stays separate from native acquisition; recording
  controls reuse the existing teleoperation owner.
• Integrate policy visualization into the platform at /#policy: built-in
  simulated sessions, submitted-target/feedback curves, inference latency and
  bounded episode replay, available through the same port in attach-only mode.
• Preserve calibrated absolute J1–J7 manual controls, trajectory/effort checks,
  queued motor reports and the calibration trajectory without encoder-zero writes.
  Integrate these with v0.8.0's official gains and acquisition observer while
  keeping pre-command observations and post-command feedback separate.
• Retain v0.8.0 demonstration collection, synchronized return-to-zero workflow,
  native recording and export; preserve read-only CAN reconnection recovery.
• Validate merged behavior with offline fake-device tests, installed-wheel tests
  and browser simulation/replay. This release adds no new physical-motion trial.

v0.8.0

• Add opt-in teleoperation demonstration collection: native acquisition sinks,
  bounded asynchronous recording, operator episode review, provenance/quality,
  read-only crash inspection, causal alignment and optional LeRobot export.
  Include integrated camera previews and offline simulated operator rehearsal.
  Complete operator-run dual-arm collection trials with three RGB cameras and
  two tactile pads, preserving native timelines and explicit episode acceptance.
• Add operator-triggered synchronized follower return to zero (J1–J7, peak
  commanded speed 5 degrees/s, gripper held), followed by relative takeover from
  the leader's current pose. No precise leader alignment or firmware-zero writes.
  Exclude recording during return/hold and retain stop/disconnect disable cleanup.
• Use official Enactic OpenArm v2 J1–J7 PD gains in shared follower hold/tracking;
  retain existing gripper gains. Operator trials validate the workflow; tracking
  error remains motion-dependent, not a full-workspace accuracy guarantee.
• Move diagnostic report writing off the CAN control loop and record dispatched
  gains, commands, pre-command observations and post-command feedback separately.

v0.7.0

• Add optional Pinocchio gravity feedforward shared by hold, panel, teleoperation
  and policy execution; explicit model/motor mappings, payload, shadow mode,
  ramp/caps/slew, firmware-range-correct torque encoding and per-cycle audit data.
  Include offline gravity CLI and analytic/lifecycle tests. Operator-authorized
  right-arm 50%/100% compensation trials completed with verified motor disable;
  full-workspace and dual-arm compensation validation remain outstanding.

• Add model-independent policy contracts, timestamped sensor histories, asynchronous
  process-isolated inference, time-anchored action chunks and independent Classic
  CAN execution/watchdogs. Include simulation, plugin CLI, recording/replay and
  offline lifecycle tests. Validate one right-arm sine cycle through the physical
  backend, including recording, gravity feedforward and verified disable.

• Simplify the standalone camera viewer with a primary head view, compact RGB
  views, responsive mobile layout, per-view status and browser update FPS.
  Preview settings and optional depth views are collapsed by default.

v0.6.0

• Add the local OpenArm Platform workspace with unchanged control, teleoperation,
  tactile and vision subpages, configured single-command startup and independent
  page connections. One loopback port proxies HTTP, streaming images, Viser
  WebSockets and both nested tactile views; remote clients need only one SSH
  forward. Motor control/teleoperation switch exclusively through the existing
  shutdown paths; external services remain externally owned. Failed sensing
  services can be retried explicitly from the workspace.
• Use NiceGUI components and its FastAPI/Uvicorn server for the platform
  dashboard, with aiohttp managing upstream HTTP and WebSocket connections.
  Add the optional platform extra; core imports and CLI help stay passive.
  Preserve mounted module views during sensing navigation and keep the existing
  motor-control lifecycle and exclusive mode switching.
• Document replacement of standalone launches, one-port SSH access and wrist
  camera recovery using distinct USB topology paths when device serials collide.
  Hardware profiles and calibration remain external to the package.
• Validate the installed platform wheel on Linux and macOS. Browser acceptance
  covers persistent module views, HTTP/video streaming, WebSockets and clean
  shutdown with software fixtures; read-only workstation checks display all four
  cameras and both tactile viewers through the platform port.

v0.5.0

• Add the optional openarm_tactile capture API and tactile extra, pinning the
  external community Paxini SDK. Core imports remain hardware-passive and free
  of GUI/serial dependencies. SDK sources and coordinate data are not vendored.
• Add openarm-runtime tactile for IP-M2324-Core 68-point serial sensors, with
  explicit device ID, independent capture, stale/error state, manual reconnect
  and bounded shutdown. Raw force counts are not calibrated newtons.
• Add a left-arm dual-sensor workspace with 3D force views, point readouts, curves
  and independent controls. Preserve the browser-confirmed sensor order; both
  sensors belong to the left arm and the right arm has no tactile sensor.
• Add explicit software simulation, external workstation configuration and
  installed-package regression coverage. Previously documented dual-sensor
  hardware trials observed pressure responses and no capture errors; release
  checks are offline and do not move motors or open tactile hardware.

v0.4.0

• Dual-arm following with independent CAN workers, all-arm preflight, coordinated
  takeover, peer-fault cancellation and serialized reports. Single-arm selection remains.

• Operator-approved voltage-only leader alarms are nonblocking during physical
  following; stale/corrupt feedback, enabled leaders and other fault bits still stop.

• Optional leader --follow adds operator-started physical following,
  relative takeover, direct MIT targets with upstream joint-bound clipping,
  65-degree gripper mapping, stop/disconnect cleanup and reports.
  Operator-run dual-arm following completed a short trial with both disables
  verified; startup remains passive. Quantified tracking accuracy is unmeasured.

• Following no longer inherits manual-slider slew/offset restrictions. Leader
  calibration and diagnostics are collapsed by default in Viser.

• Guided manual leader calibration in Viser: arm selection, stable natural-zero
  and gripper endpoint capture, visible completeness/unsaved status and persistence.

• Capture rejects moving, stale or enabled-leader feedback and retains previous
  values on failure. Advanced direction/device controls remain available.

v0.3.0

• Add independent openarm_video package and video extra, with passive imports,
  configurable RGB/OpenNI2 sources, timestamped frames, device leases, reconnect
  lifecycle, and explicit software-only simulation.
• Add openarm-runtime video, packaged four-view assets, Safari-compatible finite
  JPEG/PNG requests and raw millimetre-depth download. The vendor SDK and machine
  configuration remain outside the package.
• Fix depth/profile test path comparisons on macOS and let Linux/macOS CI jobs
  complete independently.

v0.2.0

• Final whole-arm zero convergence corrects drift after intermediate zero moves;
  all eight joints must settle together before disabled-pose checks and writes.

• Calibration moves now use slow, bounded target correction for static position
  error; actual-zero tolerance and manual-hold gains are unchanged. Failure reports
  include target/goal/feedback history captured before disable.

• Calibration move/seek speeds are now 20/10 degrees per second, with ramped
  upstream per-action Kp/Kd settings. Inactive-joint failures include the offending
  feedback.

• Single-arm v2 automatic calibration over Classic CAN, available through the
  Python API, CLI and a simplified arm/action selector in the follower panel.

• Panel assumes MIT when RID10 times out, without an advanced checkbox; explicit
  non-MIT readback is still rejected. CLI/API retain the explicit timeout option.

• Bounded stop seeking, computed-zero positioning, cancellation/disable cleanup,
  and disabled-pose checks with per-motor hardware zero write reports.

• Offline virtual-stop and failure-path tests. Both arms completed zero writes
  (right 139 s, left 126 s); all 16 angle readbacks were unchanged after one user
  power cycle. Absolute accuracy and repeatability remain unquantified.

• Existing hold/leader APIs remain compatible; native SDK and CAN bitrate are
  unchanged. See docs/calibration.md for operation and recovery.

v0.1.0

Changelog

0.1.0 — 2026-10-31

Initial public release.

• Installable core with composable leader, follower and viewer extras.
• Official OpenArm CAN1.4.1 encoding, Classic CAN telemetry, whole-arm hold and
  bounded relative targets, shared bus ownership and verified disable cleanup.
• Read-only STS3215 leader transport, background reconnect and device-bound
  software pose/gripper calibration.
• Optional Viser follower panel and live leader-to-virtual-model display.
• Public Python APIs, read-only examples, wheel-based offline CI and MIT license.

Known boundaries: Python3.12; follower hardware requires Linux and the native SDK;
GUI labels are Chinese. Physical leader-to-follower teleoperation is not included.