Issues

• Tracking: 2026-09-24 whole-ecosystem review (mongla_ws + srot-control-board + ground station + ESC flasher), against Bumblebee as the bar

  fh1m/mongla_ws#53 · Filter by author fh1m opened on Sep 24, 2026
  3

• Proposal: a runtime "shield": a few mission-independent invariants enforced on the board below the mission (max depth, no AUTO without a fresh "aided" flag, yaw-rate envelope), with a host feed that fails closed

  fh1m/mongla_ws#52 · Filter by author fh1m opened on Sep 24, 2026

• Proposal: physically based training-time augmentation (copy-paste onto empty venue frames, depth-dependent attenuation/blur, caustic overlays), which is distinct from the rejected runtime preprocessing; decide on held-out-session recall

  fh1m/mongla_ws#51 · Filter by author fh1m opened on Sep 24, 2026

• Proposal (hardware ask): a ~US$400 Ping2 echosounder as a downward altimeter, replacing the hand-typed pool_depth_m as flow's height source

  fh1m/mongla_ws#50 · Filter by author fh1m opened on Sep 24, 2026

• Proposal: run_plan "expected-value ordering" is points-per-second with no success probability; add a per-task Beta posterior counted from scorecards (VERDICT correction)

  fh1m/mongla_ws#49 · Filter by author fh1m opened on Sep 24, 2026

• Hailo compile: the measured 0.08 INT8 confidence loss is likely the DFC's reduced optimization level (320 calibration frames, likely no GPU); add a build guard that fails on "Reducing optimization level" (G-15 addendum)

  fh1m/mongla_ws#47 · Filter by author fh1m opened on Sep 24, 2026
  1

• Proposal: retrain task models on YOLOv8s (single-context on Hailo-8: 44.6 mAP @ 491 FPS vs yolov11n 39.0 @ 185; ours cap at ~98 FPS) and add yolov8s-pose keypoints → existing SQPnP; decide on held-out-session recall

  fh1m/mongla_ws#46 · Filter by author fh1m opened on Sep 24, 2026

• ESC PWM frequency (Bluejay flashed at 24 kHz, docs recommend 48 kHz) sits inside or next to the pinger bands (SAUVC 45 kHz) and the H2 ego-noise band: move to 96 kHz before any hydrophone work

  fh1m/mongla_ws#45 · Filter by author fh1m opened on Sep 24, 2026

• Proposal: undistort once at the camera (single remap in camera_node, publish rectified image + CameraInfo with D=0) to remove the distortion bug class

  fh1m/mongla_ws#42 · Filter by author fh1m opened on Sep 24, 2026

• Proposal: split auv_manager_node into a MAVLink bridge and an executive; health as DiagnosticArray; lifecycle activation; per-family actions instead of one ~60-field Move

  fh1m/mongla_ws#41 · Filter by author fh1m opened on Sep 24, 2026

• Proposal: resolve symmetric targets (bins, torpedo board) by template-rotation match counting before PnP, using the existing XFeat anchor rung

  fh1m/mongla_ws#40 · Filter by author fh1m opened on Sep 24, 2026

• Proposal: props as EKF landmarks, and any posed prop as a heading reference (re-anchor yaw globally)

  fh1m/mongla_ws#39 · Filter by author fh1m opened on Sep 24, 2026
  1

• Proposal: acoustic pinger bearing on the SROT board (3-4 hydrophones, RP2350 synchronous ADC, GCC-PHAT TDOA over MAVLink)

  fh1m/mongla_ws#38 · Filter by author fh1m opened on Sep 24, 2026
  1

• Board attitude (50 Hz, σ=0.5°) and the gyro it was computed from are fused as independent: attitude covariance collapses and the filter over-trusts the board's heading

  fh1m/mongla_ws#36 · Filter by author fh1m opened on Sep 24, 2026
  1

• A single constant pool_depth_m scales every flow velocity: sloped or stepped floors (TRANSDEC, most competition pools) give tens-of-percent scale error; add a floor-height state

  fh1m/mongla_ws#35 · Filter by author fh1m opened on Sep 24, 2026
  1

• No camera lever arm and no baro-to-camera offset: flow's v+ω×r is fused as body-origin velocity, and flow height ignores the vertical sensor offset

  fh1m/mongla_ws#27 · Filter by author fh1m opened on Sep 24, 2026
  1