Robotic grasping starts with a plain old RGB photo

What it does

Deep Object Pose Estimation (DOPE) detects specific, known objects in a scene and regresses their six-degree-of-freedom pose from an ordinary RGB camera feed. The repository ships training scripts, inference code, numerical evaluation, and synthetic data generation, plus a ROS1 Noetic node for live USB-camera inference. If you need hardware-accelerated ROS2, NVIDIA points to a separate Isaac ROS package.

The interesting bit

The repository includes synthetic data generation alongside training and inference, documented in a single walkthrough. That is the quietly useful part: it lets you manufacture labeled training scenes rather than hand-annotating real-world images.

Key highlights

• 6-DoF pose estimation for known objects using only RGB input
• Bundled support for YCB and HOPE household-object datasets with pre-trained weights
• Complete loop: synthetic data generation, training, inference, and evaluation
• Live ROS1 Noetic node for USB camera streams
• External Isaac ROS2 node available for Jetson and desktop GPU acceleration

Caveats

• The original CoRL 2018 code is archived in the repo but no longer maintained
• ROS2 inference is not included here; it requires the separate NVIDIA Isaac ROS project
• The README displays a CC BY-NC-SA 4.0 badge while the license text refers to the NVIDIA Source Code License, so the actual terms need a second look

Verdict

Grab it if you need a proven, end-to-end pipeline for known-object pose estimation in a manipulation stack. Look elsewhere if you want open-vocabulary detection or an all-in-one ROS2 solution.