Unitree Go2/Go2W · ROS 2 Humble · MuJoCo simulation + real hardware
Periodic progress reports, newest first.
Five precisions, nine scenes, two datasets on a 6 GB laptop GPU. fp16 wins outright (zero loss, +11% speed); geometry is sub-centimeter; a ground-truth audit puts honest numbers on label quality (72% YOLO, 38–62% relabeled unknowns — all at the correct position); VLFM's scorer matched and outrun 18× by MobileCLIP.
Two lines. Core navigation stack: real-robot testing separated the SLAM/planner failures into four named modes — three fixed or in hand, one under investigation. Camera line: full-rate throughput, an automatic camera-to-LiDAR calibration, and camera-LiDAR fusion wired into the SLAM; plus the semantic-mapping approach chosen (DualMap over ConceptGraphs).
Root-caused the SLAM divergence blocking real-robot runs: a missed compute deadline, not the environment as assumed. Fixed and verified — matching failures 2901 → 0. Both sensor drivers moved onboard the robot's Jetson, clearing a driver failure outstanding since July.
Unified stack with swappable components + web operator console; real-robot LiDAR occupancy navigation; RTAB-Map RGB-D instance (sim); SCAN-Planner 3D local planner (sim). Four demos, architecture diagrams, limitations, and next directions.