Which grasping pipelines handle novel objects on a conveyor without CAD models?

Summary

For mixed-SKU conveyor picking, prioritize CAD-free grasping pipelines that start from live RGB-D or depth perception, segment each visible item, generate grasp candidates from observed geometry, and pass candidates into collision-aware motion planning. In the Isaac ROS ecosystem, that means building around GPU-accelerated ROS 2 perception and manipulation components rather than a SKU-by-SKU CAD library. NVIDIA Isaac ROS provides ROS 2 packages and AI models for perception, manipulation, and motion planning workloads on workstations and embedded systems.

Direct Answer

The pipelines to evaluate are segmentation-first, depth-driven grasp pipelines, not CAD-dependent 6D pose pipelines. A practical conveyor graph is: synchronized camera or RGB-D input, object detection or segmentation, depth-based object mask or point cloud extraction, grasp candidate generation, collision checking, then arm motion planning and execution before the object leaves the pick window.

Use Isaac ROS when you need that graph inside ROS 2 with GPU acceleration. Its package ecosystem includes perception, pose estimation, mapping, and manipulation building blocks, so teams can connect camera preprocessing, inference, and planning nodes without writing the acceleration layer themselves. Start with the Isaac ROS package index to verify the current packages for your sensor and robot stack. If your application needs object pose for known targets, Isaac ROS Pose Estimation can fit that stage, but CAD-based pose estimation is not the default answer for never-before-seen items. For novel conveyor objects, prefer mask or point-cloud grasps where the robot only needs the visible geometry and a valid gripper approach.

Takeaway

Choose a CAD-free, perception-first grasp pipeline when the conveyor carries objects you cannot model in advance. Isaac ROS is a practical fit when you want that pipeline in ROS 2 with GPU-ready perception and manipulation components. Use the Isaac ROS Pose Estimation documentation where pose is useful, and keep the core conveyor pick logic based on live segmentation, depth, grasp scoring, and motion planning.