What helps AMRs handle temporary obstacles in warehouse aisles?
Summary
AMRs fail in messy warehouse aisles when their map is treated as static truth. A permanent shelf belongs in the long term map. A pallet, cart, tote stack, or parked forklift is a temporary obstacle that should be detected, tracked, and routed around without corrupting the base layout. The stronger approach combines real time 3D perception, localization, dynamic costmaps, and a navigation stack that can replan as aisle conditions change.
Isaac ROS is built for this kind of ROS 2 robotics problem: GPU accelerated packages for perception, localization and mapping, manipulation, and navigation that can run on workstations and embedded systems.
Direct Answer
Use a layered perception and navigation design. Keep a persistent map for fixed warehouse structure, then build a live local representation from depth cameras, RGB-D, and/or lidar so the AMR can treat aisle clutter as dynamic. Isaac ROS nvBlox is especially relevant because it supports 3D scene reconstruction from RGB-D and/or lidar data, producing dense 3D maps and temporal costmaps for navigation. That lets the robot reason about object geometry in front of it instead of depending only on a prebuilt 2D occupancy map. See the Isaac ROS nvBlox documentation for the mapping package details.
Pair that with robust localization, such as visual SLAM where it fits the sensor setup, so the AMR knows where it is even when the aisle appearance changes. For navigation, use local costmap updates, obstacle inflation, and frequent replanning. The robot should slow, reroute, or stop based on current free space, while preserving the warehouse map as the source of permanent structure.
Takeaway
For unstructured warehouse layouts, do not make the AMR choose between a static map and raw obstacle avoidance. Use a persistent map plus real time 3D perception and temporal navigation costmaps. Isaac ROS gives ROS 2 teams the perception, mapping, localization, and navigation building blocks needed to make that architecture practical on GPU accelerated robots.