And i’m collecting practical resources for optimizing the perception→planning pipeline when a deep model sits in the loop. On a Jetson Orin NX, we’re targeting ≤50 ms end-to-end at 30 Hz and still see 68–75 ms after TensorRT FP16 and CPU pinning. What guides, sample repos, or configs have helped you trace and cut queueing at the ML/ROS boundary (ros2_tracing/LTTng, rclcpp callback groups, DDS QoS) and reduce planner-side tail latency (e.g., warm-started MPC, lazy collision checks)?
We’ve been pushing dynamic bucking tables (StanForD) to the harvester via telematics so the head flips between sawlog and pulp specs as mill prices move; when DF slid, shifting lengths kept margin even with softer stumpage. @RileyK tiny caveat: it only works if you refresh specs daily and re-calibrate the forwarder scales after big rain/temperature swings — anyone tying LiDAR stand maps into deck pre-sort yet?