LoGoPlanner: Localization Grounded Navigation Policy with Metric-aware Visual Geometry

Jiaqi Peng  Wenzhe Cai  Yuqiang Yang  Tai Wang  Yuan Shen  Jiangmiao Pang 
Tsinghua University  Shanghai AI Laboratory 

🏡 Introduction

Most prior end-to-end navigation approaches rely on separate localization modules that require accurate sensor extrinsic calibration for self-state estimation, limiting their generalization across different robot embodiments and environments. To address this, we introduce LoGoPlanner, a localization-grounded, end-to-end navigation framework that advances the field by:

1. Finetuning a long-horizon visual-geometry backbone to ground predictions with absolute metric scale, enabling implicit state estimation for accurate localization.
2. Reconstructing surrounding scene geometry from historical observations to provide dense, fine-grained environmental awareness for reliable obstacle avoidance.
3. Conditioning the policy on implicit geometry bootstrapped by the above auxiliary tasks, thereby reducing error propagation and improving robustness.

💻 Simulation

🛠️ Installation

We use the same environment as NavDP. Please follow the installation instructions from NavDP to configure the environment:

conda activate navdp

Then install the required packages for the visual geometry model Pi3:

cd baselines/logoplanner
pip install plyfile huggingface_hub safetensors

🤔 Run the LoGoPlanner Model

Navigate to baselines/logoplanner and run the following command to start the server:

python logoplanner_server.py --port ${YOUR_PORT} --checkpoint ${SAVE_PTH_PATH}

📊 Evaluation

Open a new terminal and run the evaluation script from the {NavDP_HOME} directory:

conda activate isaaclab
python eval_startgoal_wheeled.py --port {PORT} --scene_dir {ASSET_SCENE} --scene_index {INDEX} --scene_scale {SCALE}

😉 Example

# Start the server
conda activate navdp && python logoplanner_server.py --port 19999 --checkpoint logoplanner_policy.ckpt

# Evaluate on scenes_home
conda activate isaaclab && python eval_startgoal_wheeled.py --port 19999 --scene_dir scenes_home --scene_index 0 --scene_scale 0.01

# Evaluate on cluttered_hard
conda activate isaaclab && python eval_startgoal_wheeled.py --port 19999 --scene_dir cluttered_hard --scene_index 0 --scene_scale 1.0

🤖 Real-Robot Deployment

Lekiwi is a fully open-source robotic car project developed by SIGRobotics-UIUC. It includes detailed 3D printing files and operation guides, designed to be compatible with the LeRobot imitation learning framework. It also supports the SO101 robotic arm for a complete imitation learning pipeline.

🛠️ Hardware

Compute

• Raspberry Pi 5
• Streaming to a laptop

Drive

• 3-wheel Kiwi (holonomic) drive with omni wheels

Robot Arm (Optional)

• SO-ARM101

Sensors

• RGBD camera (e.g., Intel RealSense D455)

1️⃣ 3D Printing

Parts

SIGRobotics provides ready-to-print STL files for the 3D-printed parts listed below. These can be printed with generic PLA filament on consumer-grade FDM printers. Refer to the 3D Printing section for more details.

Item	Quantity	Notes
Base plate Top	1
Base plate Bottom	1
Drive motor mount	3
Servo wheel hub	3	Requires supports1
Servo controller mount	1
12V Battery mount or 12V EU Battery mount or 5V Battery mount	1
RasPi case Top	1	2
RasPi case Bottom	1	2
Arducam base mount and wrist mount	1	Compatible with this camera
Webcam base mount, gripper insert, and wrist mount	1	Compatible with this camera
Modified Follower Arm Base	1	Use tree supports. Optional but recommended if you have not built the SO-100 arm
Follower arm	1
Leader arm	1

2️⃣ Assembly

Refer to the Assembly guide for detailed instructions.

We also recommend the following detailed tutorial from seeedstudio and its accompanying video series:

3️⃣ Installation

Install LeRobot on Raspberry Pi

1. Install Miniconda

   mkdir -p ~/miniconda3
   wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-aarch64.sh -O ~/miniconda3/miniconda.sh
   bash ~/miniconda3/miniconda.sh -b -u -p ~/miniconda3
   rm ~/miniconda3/miniconda.sh

2. Restart Shell Run source ~/.bashrc (or source ~/.bash_profile for Mac, or source ~/.zshrc for zsh).

3. Create and Activate Conda Environment

   conda create -y -n lerobot python=3.10
   conda activate lerobot

4. Clone LeRobot

   git clone https://github.com/huggingface/lerobot.git ~/lerobot

5. Install FFmpeg

   conda install ffmpeg -c conda-forge

6. Install LeRobot with LeKiwi Dependencies

   cd ~/lerobot && pip install -e ".[lekiwi]"

Install LeRobot on Laptop/PC

Follow the same steps as above for the Raspberry Pi installation.

Install RealSense SDK on Raspberry Pi

Refer to this guide.

1. Check System Version

   uname -a

2. Increase Swap Size

   sudo vim /etc/dphys-swapfile
   # Set CONF_SWAPSIZE=2048
   sudo /etc/init.d/dphys-swapfile restart
   swapon -s

3. Install Required Packages

   sudo apt-get install -y libdrm-amdgpu1 libdrm-dev libdrm-exynos1 libdrm-freedreno1 libdrm-nouveau2 libdrm-omap1 libdrm-radeon1 libdrm-tegra0 libdrm2
   sudo apt-get install -y libglu1-mesa libglu1-mesa-dev glusterfs-common libglui-dev libglui2c2
   sudo apt-get install -y mesa-utils mesa-utils-extra xorg-dev libgtk-3-dev libusb-1.0-0-dev

4. Update Udev Rules

   cd ~
   git clone https://github.com/IntelRealSense/librealsense.git
   cd librealsense
   sudo cp config/99-realsense-libusb.rules /etc/udev/rules.d/
   sudo udevadm control --reload-rules && udevadm trigger

5. Build and Install librealsense

   cd ~/librealsense
   mkdir build && cd build
   cmake .. -DBUILD_EXAMPLES=true -DCMAKE_BUILD_TYPE=Release -DFORCE_LIBUVC=true
   make -j1
   sudo make install

6. Install Python Bindings

   cd ~/librealsense/build
   cmake .. -DBUILD_PYTHON_BINDINGS=bool:true -DPYTHON_EXECUTABLE=$(which python3)
   make -j1
   sudo make install

7. Add to Python Path Edit ~/.zshrc (or your shell config file) and add:

   export PYTHONPATH=$PYTHONPATH:/usr/local/lib

   Then run source ~/.zshrc.

8. Test the Camera

   realsense-viewer

4️⃣ Motor Configuration

To identify the port for each bus servo adapter, run:

lerobot-find-port

Example output:

Finding all available ports for the MotorBus.
['/dev/ttyACM0']
Remove the USB cable from your MotorsBus and press Enter when done.

[...Disconnect the corresponding leader or follower arm and press Enter...]

The port of this MotorsBus is /dev/ttyACM0
Reconnect the USB cable.

Note: Remember to disconnect the USB cable before pressing Enter, otherwise the interface may not be detected.

On Linux, grant access to the USB ports:

sudo chmod 666 /dev/ttyACM0
sudo chmod 666 /dev/ttyACM1

Run the following command to set up the motors for LeKiwi. This will configure the arm motors (IDs 6–1) followed by the wheel motors (IDs 9, 8, 7).

lerobot-setup-motors \
    --robot.type=lekiwi \
    --robot.port=/dev/ttyACM0  # Use the port found in the previous step

5️⃣ Teleoperation

SSH into your Raspberry Pi, activate the conda environment, and run:

python -m lerobot.robots.lekiwi.lekiwi_host --robot.id=my_awesome_kiwi

On your laptop (also with the lerobot environment active), run the teleoperation example after setting the correct remote_ip and port in examples/lekiwi/teleoperate.py:

python examples/lekiwi/teleoperate.py

You should see a connection message on your laptop. You can then:

• Move the leader arm to control the follower arm.
• Use W, A, S, D to drive forward, left, backward, right.
• Use Z, X to turn left/right.
• Use R, F to increase/decrease the robot speed.

6️⃣ Deployment Preparation

Mount the RGBD camera onto LeKiwi and adjust the SO101 arm to avoid obstructing the camera view.

Tip: Before running the navigation algorithm, test the robot by having it follow simple trajectories (e.g., a sine wave or "S" curve) to ensure the MPC tracking is working correctly.

7️⃣ Deploy LoGoPlanner

On your laptop or PC, start the LoGoPlanner server:

python logoplanner_realworld_server.py --port 19999 --checkpoint ${CKPT_PATH}

Verify the server IP address:

hostname -I

On the Raspberry Pi, copy lekiwi_logoplanner_host.py to your working directory and run the client:

conda activate lerobot
python lekiwi_logoplanner_host.py --server-url http://192.168.1.100:8888 --goal-x 10 --goal-y -2

The robot will navigate to the target coordinates (10, -2). Without any external odometry module, it will use its implicit localization to reach the goal and stop.

---

Footnotes:

1: Requires 3D printing supports.
2: Raspberry Pi case parts.