# Autonomous inspection with ROS 2 and Unitree Go2 Aaron Emmanuel — independent project, October 2026. ## Development sequence 1. TurtleBot 4 / ROS 2 Humble / Gazebo Fortress established simulated LiDAR SLAM, saved-map localization, Nav2 navigation and a three-station inspection workflow. 2. OpenCV ArUco asset IDs, perspective-corrected regions and HSV classification read known indicator panels. This is controlled indicator reading, not general equipment fault diagnosis. 3. A Python ROS adapter and official C++ Unitree SDK2 worker map navigation commands into native DDS Sport Move1008 / StopMove1003 calls. Early protocol tests used a local emulator with synthetic feedback. 4. A custom native Sport API simulation server connects that SDK interface to MuJoCo motor torques, a reused ONNX walking policy and motor PD control. Measured physics state returns through SDK SportModeState and DDS to ROS. 5. An original PyTorch RGB detector adds cone/extinguisher recognition. ONNX Runtime consumes received camera pixels and publishes Detection2DArray. 6. Newly introduced stationary collision/LiDAR fixtures test replanning and explicit no-route mission disarming. The saved floorplan is unchanged. 7. Same-stamp RGB, metric depth, CameraInfo and map-to-camera TF add visible surface range and stationary-object map memory. ## Command and sensing architecture Inspection goals → Nav2 → mission-armed ROS relay / freshness gates → ROS SDK adapter → official Unitree SDK2 worker → native DDS → custom Sport simulation server → reused walking policy / PD motor torques → MuJoCo physics. Measured pose, velocity and orientation return through the SDK/DDS state channel. The Go2 profile uses simulator-truth-derived odometry. AMCL localizes against a supplied geometry-derived floorplan. This does not demonstrate Go2 SLAM or an independent hardware odometer. The simulated tilted body LiDAR casts seven rings across 360 azimuths. The raw PointCloud2 stays available; orientation leveling and a height band produce the virtual gravity-level LaserScan used by 2D navigation. The onboard RGB camera is a generic synthetic camera model, not calibrated physical Go2 sensing. ## Control correction Small gait drift at home repeatedly crossed the controller's approach/alignment boundary. A C++ plugin subclasses Humble regulated pure pursuit, enters alignment inside 0.18 m and retains it until departure exceeds 0.26 m. It retains the actual goal checker and RPP collision/regulation checks. A different goal resets phase. Recorded baseline: 16 meaningful turn-command reversals / 33.20 s alignment. Final fixed run: zero reversals / 8.44 s. This is a specific arrival regression, not general gait smoothness or precise velocity tracking. [Final arrival audit](evidence/arrival-audit.json). ## Learned detector Original eight-convolution centre/box network, 183,814 parameters, trained from scratch in PyTorch. 2,800 / 480 / 600 train / validation / test images come from 140 / 24 / 30 disjoint synthetic scenes. Checkpoint and confidence threshold were selected with validation data. No cross-split RGB hashes overlap. Independent CPU ONNX evaluation at IoU 0.50: 663 TP, 64 FP, 72 FN; 91.20% precision and 90.20% recall. False detections occurred in 15 / 148 empty test images. Views within scenes are correlated. Real-world generalization is untested. Recognition fixtures are visual-only. Learned labels do not control collision avoidance; Nav2 uses LiDAR. This detector is distinct from the reused gait policy. ## Obstacle response The crate appears only after an accepted goal and measured motion. Physical collision geometry and navigation LiDAR see it; there is no saved-map or keepout injection. In the final detour trial, the first independently received clear revised plan arrived 0.833 s after applied obstacle state. Minimum sampled padded footprint clearance was 0.439 m; named collision contacts were checked per 2 ms physics integration step. Electrical required three viewpoint refinements. The separate full-width barrier trial uses an explicit no-recovery behaviour tree. The action aborts when replanning cannot find a route; the application disarms, no later nonzero SDK publications are recorded, and physics settles. Twelve checks passed. These are nominal simulated measurements, not hardware safety or worst-case stopping bounds. They are new stationary obstacles, not moving people. [Blocked-passage audit](evidence/blocked-audit.json). ## Aligned depth and object memory Onboard RGB and MuJoCo metric depth use the same simulation state, 640×480 dimensions, stamp, optical frame and intrinsics. Depth is 32FC1 optical-Z metres; 0.1–8 m valid, NaN outside. This is ideal synthetic RGB-D, not stereo/learned depth, physical calibration or a noise model. The ROS tracker accepts the learned detection arrays, received depth/RGB, CameraInfo and stamped map-to-camera TF. Exact-stamp fusion rejects stale, invalid and missing-transform cases. Robust narrow-box surface pixels are backprojected through the intrinsics. Displayed Euclidean camera-to-visible- surface range differs from optical-Z; it is not object centre or collision clearance. Same-class one-to-one association uses 0.45 m XY / 0.30 m Z gates, withholding near ties. Three observations confirm a track. Tentative tracks expire after 1.5 s; confirmed stationary memory persists for up to 300 simulation seconds. Remembered markers fade and retain explicit age. Public numbers allocate only on confirmation and remain stable on reacquisition; private candidate IDs persist internally. This is not moving-object velocity or person tracking. ## Final integrated evidence depth_tracking_final_1: 17 / 17 mission, 8 / 8 camera, 12 / 12 obstacle and 12 / 12 depth checks. Four single-action navigation goals complete pump, valve, electrical and home. 1,720 movement publications match 1,720 SDK Move handlers; 13 Stop handlers and 3,713 DDS samples match physics. Six fixtures each obtain one distinct confirmed identity. 375 matched RGB-D/TF pairs support the tracker; delivery/fusion is not lossless. False boxes and missed detections remain in the evidence. A separate nine-frame frozen-camera opaque-occluder test passed 6 / 6 checks, preserving identity after the view changed. It is not a moving-robot or moving-person occlusion test. [Final integrated audit payloads](evidence/final-depth-audits.json). ## Recording and public numbering The final depth video is a post-run presentation replay of immutable measured data, with larger unclipped labels and public numbers 1–6. Original spectator imagery is reused; saved ROS RGB/depth are decoded and reannotated; the map is rebuilt from prior recorded events. Normal JPEG/MP4 re-encoding applies. The original live source/config snapshot and media remain archived separately. The corrected display sources have not been rerun in a new integrated live mission. Seven display checks pass, including 2,011 in-bounds map-label placements across 397 frames and nonoverlapping camera labels on all received tracking frames. [Display checks](evidence/display-tests.json). Inspection movies show separate trials at approximately 3× wall-time playback. The depth replay contains 397 frames, lasts 63.44 s and retains the original approximately 2.086 fps / 189.83 s capture cadence. Playback speed is not simulation speed. The early SDK test and [initial Go2 motion video](go2-physics-response.mp4) use normal wall-time playback. [Media hashes and original trial captions](media-provenance.json). ## Ownership and reused components Project contributions: mission and inspection software, ROS relays/interfaces, official-SDK adapter and native worker, custom simulation Sport server, C++ arrival plugin, simulated sensor publication/projection, own RGB training/inference, RGB-D fusion/stationary tracker, logging, audits and presentation tools. Upstream components: ROS 2, Nav2 planning/controllers, AMCL, SLAM Toolbox, TurtleBot 4 simulation, official Unitree SDK2, OpenCV, PyTorch/ONNX and MuJoCo. Path planning is integration of upstream Nav2 algorithms, not an original planner. Reused model/policy deployment: [wty-yy/go2_rl_gym](https://github.com/wty-yy/go2_rl_gym), commit 30e74dc507bec7a642a8c98be26081f2c6f0822d. The pre-trained walking policy was reused rather than trained for this project. The custom server is not Unitree's proprietary Sport controller. Vendor firmware modes/leases, physical sensors and calibration, hardware deployment, rough terrain, manipulation, precise velocity tracking, hard real-time operation and broader reliability remain unverified. ## Source and reproducibility The implementation, tests, trained RGB model and full raw evidence are retained locally. This website publishes selected recordings, technical notes and recorded audit results. The source archive and datasets are not publicly downloadable here. A clean second-machine full bootstrap has not been tested.