HoloBrain General Object Grasping

HoloBrain is a lightweight VLA foundation model from Horizon Robotics. It introduces an "embodied perception" architecture that fuses multi-view visual information with robot kinematics priors for stronger 3D spatial understanding and reasoning. A unified hybrid relative action space supports single-arm, dual-arm, and mobile manipulation platforms, breaking barriers across heterogeneous hardware. With only 0.2B parameters, it achieves industry-leading performance and SOTA results on benchmarks such as RoboTwin 2.0 and LIBERO, and supports complex long-horizon tasks including grasping, cloth folding, and deformable-object manipulation. Combined with the high-compute RDK S600 platform, HoloBrain enables real-time edge inference and a low-latency perception–decision–control loop, accelerating embodied AI applications. This document uses a general grasping task as an example and walks through environment setup, model deployment, and task execution of HoloBrain on RDK S600.

Platform	RDK S600 Ubuntu 24.04
ROS Version	ROS 2 Jazzy
Model	HoloBrain_v0.0_GD
Task	General object grasping
Performance (DECODE = 5)	wall=116.823ms text=9.3007ms enc=40.2941ms dec=54.6188ms

Hardware Setup

Equipment List

Device	Quantity	Image
RDK S600 development kit	1
AgileX Piper robotic arm	2
RealSense D435i	3
Robotic Arm Stand	1
Wrist Camera 3D-Printed Mount	2
Head Camera 3D-Printed Mount	1
Small Basket Prop	1

Assembly Result

Camera Check

Check USB Device Recognition

# Check whether USB devices are recognized
lsusb

Query Serial Number

# Source the ROS 2 environment before querying the serial number
source /opt/ros/jazzy/setup.bash

# Query Serial Number; this serial will be used in the HoloBrain project config launch.yaml
rs-enumerate-devices | grep Serial

Adjust the Middle Camera Position

The transform between the middle camera and the left arm is shown below. To ensure good results, adjust the middle camera so that the transform (position) is as close as possible to the figure. To determine the actual transform, see the calibration tutorial in FAQ.

Configure Robotic Arm CAN Devices

List Existing CAN Devices

# List existing CAN devices
ip link show

Rename CAN Devices

# Rename CAN devices; adjust the commands to match the actual left/right arms. The arm names should be can_left and can_right
sudo ip link set canx name can_left
sudo ip link set canx name can_right

Configure CAN Baud Rate

# After renaming the CAN devices, configure the CAN baud rate
sudo ip link set can_left type can bitrate 1000000
sudo ip link set can_right type can bitrate 1000000

Device	Purpose	Baud Rate
can_left	Left arm	1Mbps
can_right	Right arm	1Mbps

Bring Up CAN Devices

# Bring up the CAN devices for the left and right arms
sudo ifconfig can_left up
sudo ifconfig can_right up

You can verify whether the CAN devices are up with the ifconfig command:

# Check whether the CAN devices are up
ifconfig

info

The ip link configuration above is typically valid only for the current system session. After reboot, reconfigure it, or set it to start automatically via systemd.

Environment Setup

Install System Dependencies

tip

If apt reports that a package cannot be located, run sudo apt update first and then retry the install.

sudo apt update
sudo apt install tmux \
    python3-colcon-common-extensions \
    ros-jazzy-rosidl-default-generators \
    ros-jazzy-ament-cmake-auto \
    ros-jazzy-ament-lint-auto \
    ros-jazzy-ros2launch \
    ros-jazzy-foxglove-bridge \
    ros-jazzy-rosbridge-server \
    ros-jazzy-realsense2-camera

Package	Purpose
tmux	Terminal multiplexer for managing multiple background sessions
python3-colcon-common-extensions	Common colcon extensions for ROS 2 builds
ros-jazzy-rosidl-default-generators	Default generators for ROS interface definitions
ros-jazzy-ament-cmake-auto	CMake auto-configuration for the ament build system
ros-jazzy-ament-lint-auto	Auto-configuration for ament lint checks
ros-jazzy-ros2launch	ROS 2 launch file tool
ros-jazzy-foxglove-bridge	Foxglove visualization bridge
ros-jazzy-rosbridge-server	ROS 2 WebSocket bridge service
ros-jazzy-realsense2-camera	Intel RealSense depth camera driver

Clone the RoboOrchard Repository

git clone https://github.com/wunuo1/RoboOrchard

# Enter the RoboOrchard root directory
cd RoboOrchard

Create a Python Virtual Environment

Run the following commands in the RoboOrchard root directory.

# Make sure the Python version is 3.12 before creating the environment
python3 -m venv venv/roboorchard-venv

# Activate the virtual environment and load ROS 2 Jazzy
source venv/roboorchard-venv/bin/activate
source /opt/ros/jazzy/setup.bash

If everything is set up correctly, the prompt should look like:

(roboorchard-venv) root@drobot:~/VLA/holobrain/RoboOrchard#

To load them automatically, you can also add the source commands to ~/.bashrc.

Install RoboOrchard Development Dependencies

Run the following commands in the RoboOrchard root directory.

# Install RoboOrchard Python development dependencies, plus ROS 2 development, build, and related system dependencies
make dev-env
make ros2-dev-env

# Install the CPU version of PyTorch
pip install torch --index-url https://download.pytorch.org/whl/cpu

# Install the RoboOrchard application
cd python/robo_orchard_inference_app
pip install .

Build ROS 2 Packages

Run the following commands in the RoboOrchard root directory.

# Make sure the related environments are sourced
source venv/roboorchard-venv/bin/activate
source /opt/ros/jazzy/setup.bash

# Suppress Python warnings
export PYTHONWARNINGS="ignore"

# Build ROS 2 packages
make ros2-build

# Source the ROS 2 workspace after the build completes
source ros2_package/install/setup.bash

tip

If make ros2-build reports missing dependencies, run the following command first:

make ros2-dev-env

Then run again:

make ros2-build

HoloBrain Project Configuration

# Enter the HoloBrain project directory
cd projects/HoloBrain

# Install HoloBrain launch dependencies
pip install -r launch/requirements.txt

Edit the launch.yaml configuration as needed. Focus on the three camera serial numbers XXX_CAMERA_SERIAL_NO. For how to query the serial numbers, see Hardware Setup - Camera Check.

vim projects/HoloBrain/launch/launch.yaml

Directory Structure

After deployment, the main directory structure is as follows:

RoboOrchard
|-- LICENSE
|-- Makefile
|-- README.md
|-- projects
|   |-- HoloBrain
|-- pyproject.toml
|-- python
|   |-- robo_orchard_inference_app
|-- ros2_package
|   |-- Makefile
|   |-- README.md
|   |-- build
|   |-- install
|   |-- log
|   |-- requirements.txt
|   |-- robo_orchard_data_msg_ros2
|   |-- robo_orchard_data_ros2
|   |-- robo_orchard_deploy_ros2
|   |-- robo_orchard_handeye_calib_ros2
|   |-- robo_orchard_image_tools
|   |-- robo_orchard_pico_msg_ros2
|   |-- robo_orchard_piper_msg_ros2
|   |-- robo_orchard_piper_ros2
|   |-- robo_orchard_teleop_msg_ros2
|   |-- robo_orchard_teleop_ros2
|   |-- robo_orchard_wuji_glove_msg_ros2
|   |-- robo_orchard_wuji_glove_ros2
|-- scm
|   |-- lint
|   |-- qac
|   |-- requirements.txt
|-- venv
    |-- roboorchard-venv

Start/Stop HoloBrain

Start the BPU Inference Service

cd ~
wget https://sdk.d-robotics.cc/downloads/rdk_demo/rdk_s600_demo/holobrain_runtime_S600.tar.gz
tar -xvf holobrain_runtime_S600.tar.gz

# Enter the script directory
cd ./holobrain_runtime_S600/cpp_hbm_ucp/runtime/script

# Run
DECODER_STEPS=5 sh run_hbm_http_server.sh

Start holobrain_app

This script covers multiple functions such as ros_bridge, camera_service, robot_control, and inference; see launch.yaml for details.

1. Open a new terminal and run the following commands to start holobrain_app:

   # Enter RoboOrchard
   cd RoboOrchard
   
   # Make sure all environments are sourced before launch
   source venv/roboorchard-venv/bin/activate
   source /opt/ros/jazzy/setup.bash
   source ros2_package/install/setup.bash
   
   # Enter projects/HoloBrain
   cd projects/HoloBrain
   
   # Start
   ./launch/start.sh

2. After launch, you should see the following UI. Click the window numbers at the bottom to inspect each function, and make sure every function is running normally.

   

3. Open http://localhost:8501/ in a browser on the board.

4. Create a user, a task, and a description, then select the corresponding options on the right.

   

5. After making your selections, click Start to start the task, Stop to stop the task, and Reset to return the robotic arm to the zero position.

Stop holobrain_app

In the tmux terminal, press Ctrl+B, then press D to detach from the tmux session.

# Kill all processes
./launch/stop.sh