Skip to content

ml_models

Role: Asset repository for pre-trained models. No code. Just weights and exported engines, installed by colcon into share/ml_models/.

Contents

Directory What's there
yolov11_segment/ YOLOv11-small instance-segmentation .pt weights (gate, trash, slalom, floor, bins, torpedoes). Timestamped filenames (yolov11s_<task>_<YYYYMMDD>_<rev>.pt).
depth_anything/ Depth Anything v2 ViT-B ONNX models (depth_anything_v2_vitb_518.onnx, _770.onnx; ~390 MB each).

How they're consumed

  • yolo_ros_trt loads .engine files (compiled from .pt once per host via Ultralytics.YOLO.export("engine")).
  • The Depth Anything ROS node loads the .onnx and JIT-compiles a .engine on first run.

Gotchas

Hardcoded paths in configs

Vision configs reference these models by absolute path (/workspaces/isaac_ros-dev/src/ml_models/...). If the workspace is not mounted at that location, configs must be updated to match.

Engine vs onnx vs pt

The runtime loads .engine for YOLO; the engine isn't checked in. First-time setup must export .pt → .engine. For Depth Anything, configs point to .onnx.

  • No versioning beyond timestamps in filenames; no compression.