ros1

Initialize with rospy.init_node, load parameters first, create publishers before subscribers, and spin with rospy.Rate in a loop. Keep each node single-purpose rather than combining perception, planning, and control in one class.

ROS1 uses a single-threaded spinner by default, so a long-running callback blocks all others. Use rospy.MultiThreadedSpinner or offload work to a separate thread fed by a bounded queue.

Use queue_size=1 for high-frequency sensor data like images and point clouds so stale messages are dropped, and around 10 for commands you must not miss. Never use queue_size=0, which is unbounded and can cause memory bloat.

Use message_filters.ApproximateTimeSynchronizer with a slop tolerance such as 0.05 seconds instead of comparing header timestamps directly. Exact timestamp equality across different sensors almost never occurs.

Map rospy to rclpy, catkin to colcon, XML launch files to Python launch files, and nodelets to ROS2 components. Migrate leaf nodes first and use the ros1_bridge package to run both stacks during transition.

Use nodelets when nodes exchange large data like images or point clouds within the same process, since they pass shared pointers and eliminate serialization overhead. For small messages or cross-machine communication, regular nodes are sufficient.