Isaac ROS FoundationPose
Type: Model / ROS Package Tags: NVIDIA, Isaac ROS, FoundationPose, pose estimation, 6DoF, robotics, manipulation, CUDA, TensorRT, Jetson Related: NVIDIA-Isaac-ROS, NVIDIA-Isaac-for-Manipulation, Isaac-ROS-DNN-Inference, Isaac-ROS-Object-Detection, Isaac-ROS-NITROS, Isaac-ROS-cuMotion, Isaac-ROS-nvblox, NVIDIA-Isaac-Sim, NVIDIA-Jetson-Platform, TensorRT, Triton-Inference-Server Sources: https://nvidia-isaac-ros.github.io/concepts/pose_estimation/foundationpose/index.html, https://nvidia-isaac-ros.github.io/repositories_and_packages/isaac_ros_pose_estimation/isaac_ros_foundationpose/index.html, https://nvidia-isaac-ros.github.io/releases/index.html Last Updated: 2026-04-29
Summary
Isaac ROS FoundationPose is NVIDIA’s Isaac ROS support for FoundationPose, a pre-trained model for six-degree-of-freedom object pose estimation and tracking. It estimates the pose of 3D objects from visual inputs and a CAD model without retraining for each novel object, making it important for manipulation workflows that need object or goal-state pose.
Detail
Purpose
Manipulation systems need to know where target objects are in 3D before they can plan a grasp, insertion, or object-following motion. FoundationPose provides a model-based 6DoF pose estimation path that can generalize to novel objects when supplied with the required object representation and visual inputs.
Key capabilities
- 6DoF object pose estimation for 3D objects.
- Uses image/depth inputs, detection context, and a 3D CAD model rather than requiring per-object retraining.
- Tracking path for maintaining pose over time after initialization.
- Isaac ROS package and tutorials for running FoundationPose in ROS 2 pipelines.
- High GPU memory requirements for FP32 pipelines; deployment should be checked against target GPU and release notes.
- Role in NVIDIA-Isaac-for-Manipulation as a goal-state or object pose-estimation component.
NVIDIA context
FoundationPose links NVIDIA foundation-model research to practical robot manipulation. In Isaac ROS it sits between perception and planning: models estimate object state, Isaac-ROS-nvblox and sensors represent the environment, and Isaac-ROS-cuMotion plans the arm trajectory.
Connections
- NVIDIA-Isaac-ROS - parent ROS 2 package ecosystem for FoundationPose.
- NVIDIA-Isaac-for-Manipulation - reference workflow that uses pose estimation for goal-state estimation.
- Isaac-ROS-DNN-Inference - inference infrastructure adjacent to pose-estimation model execution.
- Isaac-ROS-Object-Detection - detection context can identify candidate objects before 6DoF pose estimation.
- Isaac-ROS-NITROS - accelerated image/tensor/detection transport layer for Isaac ROS perception graphs.
- Isaac-ROS-cuMotion - motion planner that can consume object or goal poses from perception.
- Isaac-ROS-nvblox - obstacle/environment representation used alongside pose estimation.
- NVIDIA-Isaac-Sim - simulation environment for generating assets, scenes, and pose-estimation tutorials.
- NVIDIA-Jetson-Platform - edge deployment target; release notes and docs should be checked for memory and performance.
- TensorRT - common optimization path for deep learning inference in Isaac ROS perception pipelines.
- Triton-Inference-Server - optional model-serving layer for complex robot perception deployments.
Source Excerpts
- NVIDIA docs describe FoundationPose as a pre-trained model for estimating 6DoF pose of 3D objects.
- The current Isaac ROS docs state that FoundationPose can handle different and novel objects without retraining.
- Isaac ROS package docs include quickstart assets and note a roughly 7 GB peak GPU memory requirement for the FP32 pipeline.