NVIDIA Isaac for Manipulation
Type: Reference Architecture Tags: NVIDIA, Isaac, Isaac ROS, robotics, manipulation, robot arms, motion planning, CUDA, cuMotion, Jetson Related: NVIDIA-Isaac, NVIDIA-Isaac-ROS, Isaac-ROS-cuMotion, Isaac-ROS-nvblox, Isaac-ROS-FoundationPose, Isaac-ROS-FoundationStereo, Isaac-ROS-DNN-Stereo-Depth, Isaac-ROS-DNN-Inference, Isaac-ROS-Object-Detection, Isaac-ROS-Image-Segmentation, Isaac-ROS-Image-Pipeline, Isaac-ROS-AprilTag, Isaac-ROS-NITROS, NVIDIA-Isaac-Sim, NVIDIA-Isaac-Lab, NVIDIA-Jetson-Platform, TensorRT Sources: https://nvidia-isaac-ros.github.io/reference_workflows/isaac_for_manipulation/reference_architecture.html, https://nvidia-isaac-ros.github.io/repositories_and_packages/isaac_ros_cumotion/index.html, https://nvidia-isaac-ros.github.io/releases/index.html Last Updated: 2026-04-29
Summary
NVIDIA Isaac for Manipulation is the current Isaac ROS reference architecture for perception-driven robot-arm manipulation. It brings together GPU-accelerated perception, depth processing, 3D reconstruction, object pose estimation, robot configuration, and Isaac-ROS-cuMotion planning so a robotic arm can perform collision-aware manipulation tasks. Current Isaac ROS release notes identify this workflow family as the current name for the workflows formerly known as Isaac Manipulator.
Detail
Purpose
Isaac for Manipulation gives robotics teams a validated starting architecture for building accelerated pick-and-place, inspection, object-following, and contact-rich manipulation systems. The workflow is designed to run with real cameras and robots, with simulated robots in NVIDIA-Isaac-Sim, and with deployment targets such as Jetson Orin or Ampere-and-newer NVIDIA GPU systems.
Key capabilities
- Reference architecture for robotic arms using NVIDIA-accelerated Isaac ROS components.
- Visual input from live or simulated RGB/depth cameras, including stereo-depth paths such as Isaac-ROS-FoundationStereo or ESS.
- Environment and obstacle perception through Isaac-ROS-nvblox scene reconstruction.
- Object and goal-state estimation with pose-estimation components such as Isaac-ROS-FoundationPose.
- DNN inference, object detection, segmentation, and stereo depth through Isaac-ROS-DNN-Inference, Isaac-ROS-Object-Detection, Isaac-ROS-Image-Segmentation, and Isaac-ROS-DNN-Stereo-Depth when workflows need object candidates or model-backed perception.
- Robot configuration through URDF plus XRDF-style collision and configuration-space metadata.
- Collision-free, time-optimized motion planning through Isaac-ROS-cuMotion and MoveIt 2 integration.
- Accelerated ROS graph transport through Isaac-ROS-NITROS.
- Sim-to-real tutorials that connect NVIDIA-Isaac-Lab, NVIDIA-Isaac-Sim, Isaac ROS, and physical robot execution.
Current naming
- Isaac ROS 4.x documentation uses Isaac for Manipulation for this current reference-workflow area.
- Earlier public material may use Isaac Manipulator; keep that as an alias in search context, but keep this page as the canonical wiki page.
- Individual tutorials, bringup packages, orchestration packages, and robot-specific examples should be folded into this page or into the relevant durable component page instead of becoming separate wiki pages.
NVIDIA context
Isaac for Manipulation is where NVIDIA’s robotics deployment stack becomes an integrated manipulation workflow. It combines NVIDIA-Isaac-ROS packages, TensorRT-optimized perception models, NVIDIA-Isaac-Sim simulation, NVIDIA-Isaac-Lab policy workflows, and NVIDIA-Jetson-Platform deployment hardware into one robot-arm architecture.
Connections
- NVIDIA-Isaac - parent robotics and physical AI platform for Isaac for Manipulation.
- NVIDIA-Isaac-ROS - ROS 2 package family that hosts the reference workflow and core components.
- Isaac-ROS-cuMotion - GPU-accelerated motion planning and robot segmentation component.
- Isaac-ROS-nvblox - 3D reconstruction and obstacle representation for collision-aware planning.
- Isaac-ROS-FoundationPose - pose-estimation model family for object/goal-state estimation.
- Isaac-ROS-FoundationStereo - stereo-depth foundation model useful when RGB stereo cameras provide depth.
- Isaac-ROS-DNN-Stereo-Depth - parent deep stereo-depth package family for ESS and FoundationStereo.
- Isaac-ROS-DNN-Inference - inference infrastructure for model-backed manipulation perception.
- Isaac-ROS-Object-Detection - detection package family for locating objects before pose estimation or manipulation.
- Isaac-ROS-Image-Segmentation - semantic segmentation package family for pixel-level object or obstacle understanding.
- Isaac-ROS-Image-Pipeline - camera preprocessing layer for manipulation perception.
- Isaac-ROS-AprilTag - fiducial detection package useful for calibration and workspace setup.
- Isaac-ROS-NITROS - accelerated transport layer for Isaac ROS manipulation graphs.
- NVIDIA-Isaac-Sim - simulation environment for manipulation tutorials, sensor simulation, and validation.
- NVIDIA-Isaac-Lab - robot-learning framework used in sim-to-real manipulation workflows.
- NVIDIA-Jetson-Platform - edge deployment family for robot-side Isaac ROS workloads.
- TensorRT - inference optimization path for perception components in the workflow.
Source Excerpts
- Current Isaac ROS docs describe Isaac for Manipulation as GPU-accelerated libraries and packages for perception-driven manipulation with robotic arms.
- The reference architecture lists visual input, environment perception, robot configuration, goal-state estimation, motion planning, and hardware platform blocks.
- Isaac ROS release notes identify Isaac for Manipulation as the current name for workflows formerly known as Isaac Manipulator.