roboreg 0.5.0

Unified eye-in-hand / eye-to-hand calibration from RGB-D images using robot mesh as calibration target.

Roboreg

Eye-to-hand calibration from RGB-D images using robot mesh as calibration target.

Mesh (purple) and Point Cloud (turqoise).

Unregistered	Registered

Table of Contents

• Installation
  • Pip (Requires CUDA Toolkit Installation)
  • Conda (Installs CUDA Toolkit)
  • Docker (Comes with CUDA Toolkit)
• Command Line Interface
  • Segment
  • Hydra Robust ICP
  • Camera Swarm
  • Monocular Differentiable Rendering
  • Stereo Differentiable Rendering
  • Render Results
• Testing

Installation

Three install options are provided:

• Pip (Requires CUDA Toolkit Installation)
• Conda (Installs CUDA Toolkit)
• Docker (Comes with CUDA Toolkit)

Pip (Requires CUDA Toolkit Installation)

[!NOTE] During runtime, CUDA Toolkit is required for the differentiable rendering. If you are planning to do differentiable rendering, see CUDA Toolkit Install Instructions. Alternatively, install using conda, see Conda (Installs CUDA Toolkit).

To pip intall roboreg, simply run

pip3 install roboreg

Conda (Installs CUDA Toolkit)

To install roboreg within an Anaconda environment (ideally Miniconda, or even better, Mamba), do the following:

1. Create an environment

   conda create -n rr-0.4.6 python=3.10
   

2. Clone this repository and install dependencies

   git clone [e-mail]:lbr-stack/roboreg.git
   mamba env update -f roboreg/env.yaml # if Anaconda or Miniconda was used, do 'conda env update -f env.yaml'
   

3. Install roboreg

   mamba activate rr-0.4.6 # can also use 'conda activate rr-0.4.6' in either case
   pip3 install roboreg/
   

Docker (Comes with CUDA Toolkit)

A sample Docker container is provided for testing purposes. First:

• Install Docker, see Docker Install Instructions
• Install NVIDIA Container Toolkit, see NVIDIA Container Toolkit Install Instructions

Next:

1. Clone this repository

   git clone [e-mail]:lbr-stack/roboreg.git
   

2. Build the Docker image (currently only runtime support, i.e. no rendering via compiled kernels)

   cd roboreg
   docker build \
       -t roboreg:latest \
       -f .docker/Dockerfile \
       --build-arg PYTORCH_VERSION=2.10.0 \
       --build-arg CUDA_VERSION=13.0 \
       .
   

   For CUDA 12.4, use:

   docker build \
       -t roboreg:latest \
       -f .docker/Dockerfile \
       --build-arg PYTORCH_VERSION=2.6.0 \
       --build-arg CUDA_VERSION=12.4 \
       .
   

3. Run container (on Linux host with NVIDIA Container Toolkit)

   docker run -it \
       --gpus all \
       --network host \
       --ipc host \
       --volume /tmp/.X11-unix:/tmp/.X11-unix \
       --volume /dev/shm:/dev/shm \
       --volume /dev:/dev --privileged \
       --env DISPLAY \
       --env QT_X11_NO_MITSHM=1 \
       roboreg:latest
   

Command Line Interface

[!TIP] Examples use sample data under test/assets/lbr_med7_r800. Data is stored via Git Large File Storage (LFS). Data is cloned automatically when git-lfs is installed. To clone in retrospect:

sudo apt install git-lfs
git lfs fetch --all
git lfs checkout

Segment

This is a required step to generate robot masks.

rr-sam2 \
    --path test/assets/lbr_med7_r800/samples \
    --pattern "left_image_*.png" \
    --n-positive-samples 5 \
    --n-negative-samples 5 \
    --device cuda

Hydra Robust ICP

The Hydra robust ICP implements a point-to-plane ICP registration on a Lie algebra. It does not use rendering and can also be used on CPU.

rr-hydra \
    --camera-info-file test/assets/lbr_med7_r800/samples/left_camera_info.yaml \
    --path test/assets/lbr_med7_r800/samples \
    --mask-pattern mask_sam2_left_image_*.png \
    --depth-pattern depth_*.npy \
    --joint-states-pattern joint_states_*.npy \
    --urdf-path test/assets/lbr_med7_r800/description/lbr_med7_r800.urdf \
    --root-link-name lbr_link_0 \
    --end-link-name lbr_link_7 \
    --number-of-points 5000 \
    --output-file HT_hydra_robust.npy

Camera Swarm

[!WARNING] On first run, nvdiffrast compiles PyTorch extensions. This might use too many resources on some systems (< 16 GB RAM). You can create an environment variable export MAX_JOBS=1 before the first run to limit concurrent compilation. Also refer to this Issue.

The camera swarm optimization can serve for finding an initial guess to Monocular Differentiable Rendering or Stereo Differentiable Rendering.

rr-cam-swarm \
    --collision-meshes \
    --n-cameras 1000 \
    --min-distance 0.5 \
    --max-distance 3.0 \
    --angle-range 3.141 \
    --w 0.7 \
    --c1 1.5 \
    --c2 1.5 \
    --max-iterations 100 \
    --display-progress \
    --urdf-path test/assets/lbr_med7_r800/description/lbr_med7_r800.urdf \
    --root-link-name lbr_link_0 \
    --end-link-name lbr_link_7 \
    --target-reduction 0.8 \
    --scale 0.1 \
    --n-samples 1 \
    --camera-info-file test/assets/lbr_med7_r800/samples/left_camera_info.yaml \
    --path test/assets/lbr_med7_r800/samples \
    --image-pattern left_image_*.png \
    --joint-states-pattern joint_states_*.npy \
    --mask-pattern mask_sam2_left_image_*.png \
    --output-file HT_cam_swarm.npy

Monocular Differentiable Rendering

[!WARNING] On first run, nvdiffrast compiles PyTorch extensions. This might use too many resources on some systems (< 16 GB RAM). You can create an environment variable export MAX_JOBS=1 before the first run to limit concurrent compilation. Also refer to this Issue.

This monocular differentiable rendering refinement requires a good initial estimate, as e.g. obtained from Hydra Robust ICP or Camera Swarm

rr-mono-dr \
    --optimizer SGD \
    --lr 0.01 \
    --max-iterations 100 \
    --display-progress \
    --urdf-path test/assets/lbr_med7_r800/description/lbr_med7_r800.urdf \
    --root-link-name lbr_link_0 \
    --end-link-name lbr_link_7 \
    --camera-info-file test/assets/lbr_med7_r800/samples/left_camera_info.yaml \
    --extrinsics-file test/assets/lbr_med7_r800/samples/HT_hydra_robust.npy \
    --path test/assets/lbr_med7_r800/samples \
    --image-pattern left_image_*.png \
    --joint-states-pattern joint_states_*.npy \
    --mask-pattern mask_sam2_left_image_*.png \
    --output-file HT_dr.npy

Stereo Differentiable Rendering

[!WARNING] On first run, nvdiffrast compiles PyTorch extensions. This might use too many resources on some systems (< 16 GB RAM). You can create an environment variable export MAX_JOBS=1 before the first run to limit concurrent compilation. Also refer to this Issue.

This stereo differentiable rendering refinement requires a good initial estimate, as e.g. obtained from Hydra Robust ICP or Camera Swarm

rr-stereo-dr \
    --optimizer SGD \
    --lr 0.01 \
    --max-iterations 100 \
    --display-progress \
    --urdf-path test/assets/lbr_med7_r800/description/lbr_med7_r800.urdf \
    --root-link-name lbr_link_0 \
    --end-link-name lbr_link_7 \
    --left-camera-info-file test/assets/lbr_med7_r800/samples/left_camera_info.yaml \
    --right-camera-info-file test/assets/lbr_med7_r800/samples/right_camera_info.yaml \
    --left-extrinsics-file test/assets/lbr_med7_r800/samples/HT_hydra_robust.npy \
    --right-extrinsics-file test/assets/lbr_med7_r800/samples/HT_right_to_left.npy \
    --path test/assets/lbr_med7_r800/samples \
    --left-image-pattern left_image_*.png \
    --right-image-pattern right_image_*.png \
    --joint-states-pattern joint_states_*.npy \
    --left-mask-pattern mask_sam2_left_image_*.png \
    --right-mask-pattern mask_sam2_right_image_*.png \
    --left-output-file HT_left_dr.npy \
    --right-output-file HT_right_dr.npy

Render Results

[!WARNING] On first run, nvdiffrast compiles PyTorch extensions. This might use too many resources on some systems (< 16 GB RAM). You can create an environment variable export MAX_JOBS=1 before the first run to limit concurrent compilation. Also refer to this Issue.

Generate renders using the obtained extrinsics:

rr-render \
    --batch-size 1 \
    --num-workers 0 \
    --urdf-path test/assets/lbr_med7_r800/description/lbr_med7_r800.urdf \
    --root-link-name lbr_link_0 \
    --end-link-name lbr_link_7 \
    --camera-info-file test/assets/lbr_med7_r800/samples/left_camera_info.yaml \
    --extrinsics-file test/assets/lbr_med7_r800/samples/HT_left_dr.npy \
    --images-path test/assets/lbr_med7_r800/samples \
    --joint-states-path test/assets/lbr_med7_r800/samples \
    --image-pattern left_image_*.png \
    --joint-states-pattern joint_states_*.npy \
    --output-path /tmp/renders/lbr_med7_r800

Testing

For testing on the xarm data, follow Docker (Comes with CUDA Toolkit). Inside the container, do

Hydra Robust ICP

To run Hydra robust ICP on provided xarm and realsense data, run

rr-hydra \
    --camera-info-file test/assets/xarm_7/samples/camera_info.yaml \
    --path test/assets/xarm_7/samples \
    --mask-pattern mask_*.png \
    --depth-pattern depth_*.npy \
    --joint-states-pattern joint_state_*.npy \
    --urdf-path test/assets/xarm_7/description/xarm_7.urdf \
    --root-link-name link_base \
    --end-link-name link7 \
    --number-of-points 5000 \
    --output-file HT_hydra_robust.npy

Render Results

Generate renders using the obtained extrinsics:

rr-render \
    --batch-size 1 \
    --num-workers 0 \
    --urdf-path test/assets/xarm_7/description/xarm_7.urdf \
    --root-link-name link_base \
    --end-link-name link7 \
    --camera-info-file test/assets/xarm_7/samples/camera_info.yaml \
    --extrinsics-file test/assets/xarm_7/samples/HT_hydra_robust.npy \
    --images-path test/assets/xarm_7/samples \
    --joint-states-path test/assets/xarm_7/samples \
    --image-pattern img_*.png \
    --joint-states-pattern joint_state_*.npy \
    --output-path /tmp/renders/xarm_7

Acknowledgements

Organizations and Grants

We would further like to acknowledge following supporters:

Logo	Notes
	This work was supported by core and project funding from the Wellcome/EPSRC [WT203148/Z/16/Z; NS/A000049/1; WT101957; NS/A000027/1].
	This project has received funding from the European Union's Horizon 2020 research and innovation programme under grant agreement No 101016985 (FAROS project).
	Built at RViMLab.
	Built at CAI4CAI.
	Built at King's College London.