OpenArmX VLA User Guide (LeRobot)

This document explains how to use the OpenArmX robot with the LeRobot framework for VLA data collection, ACT training, and inference.

✅ Success Tips ⚠️ Notes 🚨 High-Risk Items 💡 Practical Advice

🚨 The 3 most important things in this document:
1) The startup order must be followed strictly step by step.
2) W/H/FPS must be exactly the same across camera publishing, collection, and inference.
3) Before collection, update dataset parameters in config/vla_collect.env.

Table of Contents

1. Device Roles
2. General Prerequisites
3. VLA Data Collection Workflow (IPC)
   • 3.1 Manual Startup Order (must follow order)
   • 3.2 GUI One-Click Startup (recommended)
   • 3.3 Quick Topic Check (optional)
   • 3.4 Recording UI Key Guide
   • 3.5 Data Collection Command Parameters
4. ACT Training Workflow (User High-Performance PC/Server)
   • 4.1 Notes Before Training
   • 4.2 Download ACT Dependency Models
   • 4.3 ACT Training Commands
5. VLA Inference Workflow (IPC + Inference Machine)
   • 5.1 Inference Prerequisites
   • 5.2 Startup Order
   • 5.3 Inference Command Parameters
6. ROS_DOMAIN_ID Configuration for Two-Machine Collaboration
   • 6.1 Check Current Configuration
   • 6.2 Set Both Machines to the Same Value (e.g., 77, if inconsistent)
   • 6.3 Verify Again
7. Camera Parameter Configuration Reference
   • 7.1 Which Parameters Need to Be Modified in Commands
   • 7.2 Available Resolution/FPS Combinations (D405 / D435)
   • 7.3 Three-Camera Bandwidth Limit and Recommended Settings

---

🧩 1. Device Roles

• IPC (provided by us): Robot CAN control, Pico VR teleoperation, 3-camera publishing, LeRobot data collection.
• User machine (self-configured): Model training and inference (can work together with the IPC).

✅ 2. General Prerequisites

• The IPC workspace has been built, and source ~/openarmx_ws/install/setup.bash works properly.
• The robot can start normally and can be teleoperated through VR.
• Two-machine communication is required for the inference scenario in Section 5; configure DOMAIN ID according to Section 6.

⚠️ If Section 2 is not satisfied, the following steps will likely fail.

---

📦 3. VLA Data Collection Workflow (IPC)

🚦 3.1 Manual Startup Order (must follow order)

🚨 Do not skip steps or start modules in parallel out of order.

Step 1: Start the Real Robot

cd ~/openarmx_ws
source install/setup.bash
ros2 launch openarmx_bringup openarmx.bimanual.launch.py \
  control_mode:=mit \
  robot_controller:=forward_position_controller \
  use_fake_hardware:=false

Step 2: Start Pico Bridge

cd ~/openarmx_ws
source install/setup.bash
ros2 run openarmx_teleop_bridge_vr_pico openarmx_teleop_bridge_vr_pico_node

Step 3: Start VR Teleoperation

cd ~/openarmx_ws
source install/setup.bash
ros2 launch openarmx_teleop_vr_pico teleop_vr_pico.launch.py

Step 4: Start Three-Camera Publishing

First replace the camera model and serial number in the command with your own device parameters:

• Supported camera models: D435, D405
• cam_left_* / cam_right_* / cam_head_* correspond to left hand, right hand, and head camera respectively
• Query serial numbers: rs-enumerate-devices | grep "Serial Number"

With the standard IPC + standard docking station, the stable upper limit for three cameras is 640x480 @ 30fps. For camera parameter selection and available combinations, see Section 7 "Camera Parameter Configuration Reference".

💡 Recommended default: run through the full pipeline with 424x240 @ 30fps first, then gradually increase resolution.

cd ~/openarmx_ws
source install/setup.bash
W=424; H=240; FPS=30
ros2 launch openarmx_lerobot camera_publisher.launch.py \
  width:=$W height:=$H fps:=$FPS \
  cam_left_serial:=218622270388 cam_left_type:=D405 \
  cam_right_serial:=218622274446 cam_right_type:=D405 \
  cam_head_serial:=335522070220 cam_head_type:=D435

You need to modify these parameters based on your actual devices:

• W / H / FPS: unified resolution and frame rate for all three cameras (example: 424x240@30).
• cam_left_serial / cam_right_serial / cam_head_serial: replace with your three camera serial numbers.
• cam_left_type / cam_right_type / cam_head_type: set to actual camera model D405 or D435.

If you need to tune exposure parameters for all three cameras at startup, you can use the following example:

cd ~/openarmx_ws
source install/setup.bash
W=424; H=240; FPS=30
ros2 launch openarmx_lerobot camera_publisher.launch.py \
  width:=$W height:=$H fps:=$FPS \
  cam_left_serial:=218622270388 cam_left_type:=D405 \
  cam_right_serial:=218622274446 cam_right_type:=D405 \
  cam_head_serial:=335522070220 cam_head_type:=D435 \
  cam_left_color_auto_exposure:=true \
  cam_left_color_exposure:=10000 \
  cam_left_color_gain:=32 \
  cam_right_color_auto_exposure:=true \
  cam_right_color_exposure:=10000 \
  cam_right_color_gain:=32 \
  cam_head_color_auto_exposure:=true \
  cam_head_color_exposure:=10000 \
  cam_head_color_gain:=16

Common adjustable color parameters:

• cam_*_color_auto_exposure: color auto exposure, values true/false/unset
• cam_*_color_exposure: color manual exposure, range 1..10000
• cam_*_color_gain: color manual gain, range 0..128
• cam_*_color_auto_white_balance: color auto white balance, values true/false/unset
• cam_*_color_white_balance: color manual white balance, range 2800..6500
• cam_*_color_brightness: brightness, range -64..64
• cam_*_color_contrast: contrast, range 0..100
• cam_*_color_saturation: saturation, range 0..100
• cam_*_color_sharpness: sharpness, range 0..100

Notes:

• cam_left_* / cam_right_* / cam_head_* apply to left hand, right hand, and head camera respectively
• unset means do not force-set this parameter and keep default driver behavior
• If only cam_*_color_exposure or cam_*_color_gain is specified, launch will automatically add cam_*_color_auto_exposure:=false
• If only cam_*_color_white_balance is specified, launch will automatically add cam_*_color_auto_white_balance:=false

Step 5: Start LeRobot Data Collection

Enter the LeRobot environment first, then run the recording command:

• W/H/FPS configures camera resolution and frame rate during data collection (for example: W=640; H=480; FPS=30).
• Here, W/H/FPS must be exactly the same as width/height/fps in camera_publisher.launch.py.
• After changing W/H/FPS in camera publisher, update W/H/FPS in the collection command accordingly; otherwise format mismatch will cause errors.

🚨 Key constraint: Collection W/H/FPS = Camera publish width/height/fps.

lerobot-env
W=424; H=240; FPS=30
HF_HUB_OFFLINE=1 lerobot-record \
  --robot.type=openarmx_follower_ros2 \
  --robot.cameras="{cam_left: {type: ros2, image_topic: /cam_left/color/image, depth_topic: /cam_left/depth/image, use_depth: true, width: $W, height: $H, fps: $FPS}, cam_right: {type: ros2, image_topic: /cam_right/color/image, depth_topic: /cam_right/depth/image, use_depth: true, width: $W, height: $H, fps: $FPS}, cam_head: {type: ros2, image_topic: /cam_head/color/image, depth_topic: /cam_head/depth/image, use_depth: true, width: $W, height: $H, fps: $FPS}}" \
  --teleop.type=openarmx_leader_ros2 \
  --dataset.repo_id=local/your_dataset_name \
  --dataset.single_task="task_name_you_perform" \
  --dataset.num_episodes=total_number_of_episodes \
  --dataset.episode_time_s=duration_per_episode_seconds \
  --dataset.reset_time_s=interval_after_each_episode \
  --dataset.push_to_hub=false \
  --display_data=true

Example:

lerobot-env
W=424; H=240; FPS=30
HF_HUB_OFFLINE=1 lerobot-record \
  --robot.type=openarmx_follower_ros2 \
  --robot.cameras="{cam_left: {type: ros2, image_topic: /cam_left/color/image, depth_topic: /cam_left/depth/image, use_depth: true, width: $W, height: $H, fps: $FPS}, cam_right: {type: ros2, image_topic: /cam_right/color/image, depth_topic: /cam_right/depth/image, use_depth: true, width: $W, height: $H, fps: $FPS}, cam_head: {type: ros2, image_topic: /cam_head/color/image, depth_topic: /cam_head/depth/image, use_depth: true, width: $W, height: $H, fps: $FPS}}" \
  --teleop.type=openarmx_leader_ros2 \
  --dataset.repo_id=local/take_box \
  --dataset.single_task="take box" \
  --dataset.num_episodes=70 \
  --dataset.episode_time_s=180 \
  --dataset.reset_time_s=5 \
  --dataset.push_to_hub=false \
  --display_data=true

🚀 3.2 GUI One-Click Startup (recommended)

If you want to automatically launch multiple terminal windows in a fixed order, like Section 3.1, you can use the one-click startup script in this repository:

• scripts/vla_collect_gui.sh: GUI multi-terminal one-click startup script
• ⚠️ config/vla_collect.env: one-click startup config file, centrally storing robot/camera/data-collection parameters. Please prioritize parameter changes here
• scripts/README_GUI_EN.md: standalone guide for one-click startup

Recommended to enter the repository directory first:

cd /home/openarmx/openarmx_ws/src/openarmx_vla

Run a pre-check before startup:

bash scripts/vla_collect_gui.sh check

The pre-check validates:

• Whether WORKSPACE_DIR/install/setup.bash exists
• Whether current session is a graphical desktop (DISPLAY / WAYLAND_DISPLAY)
• Whether terminal command specified by GUI_TERMINAL is available (default gnome-terminal)
• Whether ros2 is available
• In collect mode, whether LEROBOT_ENV_CMD (default lerobot-env) is found in interactive shell

Common startup modes:

# 1. Start only real robot + Pico Bridge + VR Teleop
bash scripts/vla_collect_gui.sh base

# 2. Start robot base stack + camera publishing
bash scripts/vla_collect_gui.sh base_camera

# 3. Start robot base stack + camera publishing + LeRobot data collection
# ⚠️ Note: before each one-click collection run, update DATASET_REPO_ID in config/vla_collect.env
bash scripts/vla_collect_gui.sh collect

# Close all terminals started by this script
bash scripts/vla_collect_gui.sh stop

⚠️ Always check DATASET_REPO_ID before collect to avoid writing to the wrong dataset directory.

Mode mapping:

• base: equivalent to manually running "real robot + Pico Bridge + VR teleop"
• base_camera: starts three-camera publishing on top of base
• collect: starts LeRobot data collection on top of base_camera
• stop: precisely closes windows launched by this script via state file; no error even if some windows were closed manually

Script startup behavior:

• Pops up multiple terminal windows in sequence and starts each module with configured delays
• Each window remains open after command execution for on-site troubleshooting
• If windows from the previous run are still active, it prompts you to run bash scripts/vla_collect_gui.sh stop first
• In collect mode, it runs LEROBOT_ENV_CMD first in recording terminal, then runs lerobot-record
• Default config path is config/vla_collect.env; you can also temporarily switch with VLA_CONFIG_FILE=/your_path.env bash scripts/vla_collect_gui.sh collect

Usually, you only need to modify these items in config/vla_collect.env:

• Base path and GUI parameters: workspace path, terminal command, window state file path
• Robot base parameters: CONTROL_MODE, ROBOT_CONTROLLER, USE_FAKE_HARDWARE
• VR teleoperation parameters: Pico/Teleop control rate, grasp threshold, topic names, etc.
• Camera parameters: W/H/FPS, three camera serial numbers, camera model, exposure/gain/white balance, etc.
• Data collection parameters: dataset name, task description, episode count, episode duration, reset time, whether to display data, etc.

Special attention:

• W/H/FPS is used by both camera publishing and lerobot-record; keep it consistent with actual camera output
• CAM_LEFT_TYPE / CAM_RIGHT_TYPE / CAM_HEAD_TYPE must be set to actual devices D405 or D435
• CAM_LEFT_SERIAL / CAM_RIGHT_SERIAL / CAM_HEAD_SERIAL must be replaced with your camera serial numbers
• If using collect mode, make sure dataset parameters like DATASET_REPO_ID and DATASET_SINGLE_TASK are updated to your task

🚨 The 4 items above are frequent failure points; check them one by one before each collection run.

For a more complete one-click startup guide, refer to: scripts/README_GUI_EN.md.

🔎 3.3 Quick Topic Check (optional)

ros2 topic list | grep cam
ros2 topic list | grep joint_states
ros2 topic list | grep forward_position_controller/commands

Expected at minimum:

• Camera topics: /cam_left/color/image, /cam_right/color/image, /cam_head/color/image
• Joint state: /joint_states
• Teleop outputs: /left_forward_position_controller/commands, /right_forward_position_controller/commands

If the above conditions are met, data collection can proceed.

⌨️ 3.4 Recording UI Key Guide

• → (Right Arrow): End and save current episode, then enter reset stage.
• ← (Left Arrow): Discard current episode and re-record.
• Esc: Stop recording and exit, then save dataset.

🧾 3.5 Data Collection Command Parameters

Common parameters:

• HF_HUB_OFFLINE=1: Enable Hugging Face Hub offline mode.
• --robot.type=openarmx_follower_ros2: Specify target robot type being controlled.
• --teleop.type=openarmx_leader_ros2: Specify teleoperation device type.
• --dataset.repo_id=local/xxx: Dataset storage identifier (path under ~/.cache/huggingface/lerobot/local/).
• --dataset.single_task: Task description.
• --dataset.num_episodes: Total number of episodes.
• --dataset.episode_time_s: Maximum duration per episode (seconds).
• --dataset.reset_time_s: Reset wait time between episodes (seconds).
• --dataset.push_to_hub: Whether to upload to Hugging Face Hub.
• --display_data: Whether to display real-time data.

Other parameters:

• --dataset.root: Custom dataset save path.
• --dataset.fps: Limit recording frame rate.
• --dataset.video: Whether to encode images as video.
• --dataset.vcodec: Video codec (default libsvtav1).
• --dataset.video_encoding_batch_size: Number of episodes per batch video encoding.
• --dataset.private: Set private when uploading to Hub.
• --dataset.tags: Hub dataset tags.
• --dataset.num_image_writer_processes: Number of image writer processes.
• --dataset.num_image_writer_threads_per_camera: Number of writer threads per camera.
• --dataset.rename_map: Rename observation keys.

---

🧠 4. ACT Training Workflow (User High-Performance PC/Server)

📌 4.1 Notes Before Training

ACT training is recommended on a user-provided high-performance PC or server (dedicated GPU recommended). Before training, complete LeRobot environment installation on this machine and download required model files. This document uses ACT as an example. For more environment setup and model training tutorials, see the official docs: http://docs.openarmx.com/.

📥 4.2 Download ACT Dependency Models

Run in a LeRobot environment terminal:

mkdir -p ~/.cache/torch/hub/checkpoints
# Enter LeRobot environment and install dependencies
lerobot-env
wget https://mirrors.tuna.tsinghua.edu.cn/pytorch/models/resnet18-f37072fd.pth \
  -O ~/.cache/torch/hub/checkpoints/resnet18-f37072fd.pth

🏋️ 4.3 ACT Training Commands

4.3.1 Single-GPU Training (optional)

lerobot-env
export HF_HUB_OFFLINE=1
export TRANSFORMERS_OFFLINE=1
lerobot-train \
  --dataset.repo_id=local/your_dataset_name \
  --dataset.root=absolute_path_to_your_dataset \
  --policy.type=act \
  --policy.push_to_hub=false \
  --output_dir=outputs/your_trained_model_name \
  --batch_size=batch_size_per_training_step \
  --steps=total_training_steps \
  --log_freq=log_every_n_steps \
  --save_freq=save_every_n_steps

4.3.2 Multi-GPU Training (optional)

lerobot-env
export HF_HUB_OFFLINE=1
export TRANSFORMERS_OFFLINE=1
torchrun --nproc_per_node=number_of_your_gpus \
  "$(which lerobot-train)" \
  --dataset.repo_id=local/your_dataset_name \
  --dataset.root=absolute_path_to_your_dataset \
  --policy.type=act \
  --policy.push_to_hub=false \
  --output_dir=outputs/your_trained_model_name \
  --batch_size=batch_size_per_training_step \
  --steps=total_training_steps \
  --log_freq=log_every_n_steps \
  --save_freq=save_every_n_steps

After training, record the exported pretrained_model path in output_dir for inference in Section 5.

---

🤖 5. VLA Inference Workflow (IPC + Inference Machine)

Using ACT as an example, this section explains how to load a trained model for online inference. In the current workflow, two-machine communication between IPC and user machine is required. Complete ROS_DOMAIN_ID configuration in Section 6 before inference.

🚨 If two-machine communication is not configured before inference, linkage will almost certainly fail.

✅ 5.1 Inference Prerequisites

• Training is completed and model path is available (usually the pretrained_model directory).
• IPC can start robot and cameras normally.
• Inference is usually run on another user machine, so two-machine communication is required; complete ROS_DOMAIN_ID configuration in Section 6 first.

🚦 5.2 Startup Order

Step 1: Start Real Robot (IPC)

cd ~/openarmx_ws
source install/setup.bash
ros2 launch openarmx_bringup openarmx.bimanual.launch.py \
  control_mode:=mit \
  robot_controller:=forward_position_controller \
  use_fake_hardware:=false

Step 2: Start Three-Camera Publishing (IPC)

Modify W/H/FPS and the three camera serial/type according to Section 7 "Camera Parameter Configuration Reference".

cd ~/openarmx_ws
source install/setup.bash
W=424; H=240; FPS=30
ros2 launch openarmx_lerobot camera_publisher.launch.py \
  width:=$W height:=$H fps:=$FPS \
  cam_left_serial:=218622270388 cam_left_type:=D405 \
  cam_right_serial:=218622274446 cam_right_type:=D405 \
  cam_head_serial:=335522070220 cam_head_type:=D435

If you need fixed exposure for all three cameras before inference, you can append the same three-camera exposure settings used in Section 3:

• Left hand: cam_left_color_auto_exposure:=false cam_left_color_exposure:=400 cam_left_color_gain:=32
• Right hand: cam_right_color_auto_exposure:=false cam_right_color_exposure:=400 cam_right_color_gain:=32
• Head: cam_head_color_auto_exposure:=false cam_head_color_exposure:=300 cam_head_color_gain:=16

Step 3: Run Inference (Inference Machine)

• W/H/FPS in the inference command must be exactly the same as width/height/fps of the current camera publishing node.
• During inference, camera format (resolution/frame rate) should match data collection format (recommended: same format as training data for this model).

🚨 Key constraint: Inference W/H/FPS = Collection W/H/FPS = Camera publish width/height/fps.

lerobot-env
W=424; H=240; FPS=30
HF_HUB_OFFLINE=1 lerobot-record \
  --robot.type=openarmx_follower_ros2 \
  --robot.cameras="{cam_left: {type: ros2, image_topic: /cam_left/color/image, depth_topic: /cam_left/depth/image, use_depth: true, width: $W, height: $H, fps: $FPS}, cam_right: {type: ros2, image_topic: /cam_right/color/image, depth_topic: /cam_right/depth/image, use_depth: true, width: $W, height: $H, fps: $FPS}, cam_head: {type: ros2, image_topic: /cam_head/color/image, depth_topic: /cam_head/depth/image, use_depth: true, width: $W, height: $H, fps: $FPS}}" \
  --robot.skip_send_action=false \
  --dataset.repo_id="local/inference_result_model_name" \
  --dataset.single_task="your_task_name" \
  --dataset.num_episodes=number_of_inference_runs \
  --dataset.push_to_hub=false \
  --display_data=true \
  --policy.path="path_to_your_trained_model"

Example:

lerobot-env
W=424; H=240; FPS=30
HF_HUB_OFFLINE=1 lerobot-record \
  --robot.type=openarmx_follower_ros2 \
  --robot.cameras="{cam_left: {type: ros2, image_topic: /cam_left/color/image, depth_topic: /cam_left/depth/image, use_depth: true, width: $W, height: $H, fps: $FPS}, cam_right: {type: ros2, image_topic: /cam_right/color/image, depth_topic: /cam_right/depth/image, use_depth: true, width: $W, height: $H, fps: $FPS}, cam_head: {type: ros2, image_topic: /cam_head/color/image, depth_topic: /cam_head/depth/image, use_depth: true, width: $W, height: $H, fps: $FPS}}" \
  --robot.skip_send_action=false \
  --dataset.repo_id=local/eval_take_box \
  --dataset.single_task="take the box" \
  --dataset.num_episodes=10 \
  --dataset.push_to_hub=false \
  --display_data=true \
  --policy.path="/home/i4090/openarmx_vla/src/VLA/OUTPUTS/045000/pretrained_model"

🧾 5.3 Inference Command Parameters

• HF_HUB_OFFLINE=1: Offline mode, do not fetch online resources.
• --robot.type=openarmx_follower_ros2: Target robot type for inference action publishing.
• --robot.skip_send_action=false: false means send real actions; true means validate pipeline only without moving robot.
• --dataset.repo_id="local/inference_result_model_name": Identifier for inference result storage.
• --dataset.single_task="your_task_name": Inference task name (metadata).
• --dataset.num_episodes=number_of_inference_runs: Number of inference episodes.
• --dataset.push_to_hub=false: Do not upload to Hugging Face Hub.
• --display_data=true: Display inference process data.
• --policy.path="path_to_your_trained_model": Local model path.

---

🌐 6. ROS_DOMAIN_ID Configuration for Two-Machine Collaboration

When IPC and user machine need cross-machine communication, ROS_DOMAIN_ID must be identical on both machines (recommended: same value, e.g. 77).

🚨 If ROS_DOMAIN_ID is inconsistent between the two machines, cross-machine topic discovery will fail.

6.1 Check Current Configuration

Run on both machines separately:

echo $ROS_DOMAIN_ID

6.2 Set Both Machines to the Same Value (e.g., 77, if inconsistent)

Run on both machines (set DOMAIN_ID to a shared value first, here 77):

DOMAIN_ID=77
grep -q '^export ROS_DOMAIN_ID=' ~/.bashrc \
  && sed -i "s/^export ROS_DOMAIN_ID=.*/export ROS_DOMAIN_ID=${DOMAIN_ID}/" ~/.bashrc \
  || echo "export ROS_DOMAIN_ID=${DOMAIN_ID}" >> ~/.bashrc
source ~/.bashrc

6.3 Verify Again

echo $ROS_DOMAIN_ID

Both machines should print the same value.

---

📷 7. Camera Parameter Configuration Reference

🛠️ 7.1 Which Parameters Need to Be Modified in Commands

When using camera_publisher.launch.py, usually you only need to modify these parameters:

• W / H / FPS: resolution and frame rate.
• cam_left_serial / cam_right_serial / cam_head_serial: serial numbers of the three cameras.
• cam_left_type / cam_right_type / cam_head_type: camera model (D405 or D435) for each camera.
• In the parameter names below, * is not a literal character but a placeholder and must be replaced with a specific camera prefix:
  • Use cam_left for left hand camera
  • Use cam_right for right hand camera
  • Use cam_head for head camera
• For example, cam_*_color_exposure should actually be cam_left_color_exposure, cam_right_color_exposure, or cam_head_color_exposure.
• cam_*_color_auto_exposure: color auto exposure, values true/false/unset.
• cam_*_color_exposure: color manual exposure, range 1..10000.
• cam_*_color_gain: color manual gain, range 0..128.
• cam_*_color_auto_white_balance: color auto white balance, values true/false/unset.
• cam_*_color_white_balance: color manual white balance, range 2800..6500.
• cam_*_color_brightness: brightness, range -64..64.
• cam_*_color_contrast: contrast, range 0..100.
• cam_*_color_saturation: saturation, range 0..100.
• cam_*_color_sharpness: sharpness, range 0..100.

The same W/H/FPS must also be set in lerobot-record commands (collection and inference), and must stay consistent:

• W/H/FPS in lerobot-record = width/height/fps in camera_publisher.launch.py.
• Inference W/H/FPS = Data collection W/H/FPS (recommended to match the training data format used by the model).

🚨 Consistency here is the core stability constraint of the entire pipeline.

Serial number query command:

rs-enumerate-devices | grep "Serial Number"

📐 7.2 Available Resolution/FPS Combinations (D405 / D435)

camera_publisher.launch.py has built-in validation; only the following valid combinations can be used:

Intel RealSense D405

Resolution	Supported FPS
1280 x 720	5, 15, 30
848 x 480	5, 15, 30, 60, 90
640 x 480	5, 15, 30, 60, 90
640 x 360	5, 15, 30, 60, 90
480 x 270	5, 15, 30, 60, 90
424 x 240	5, 15, 30, 60, 90

Intel RealSense D435 / D435i

Resolution	Supported FPS
1920 x 1080	6, 15, 30
1280 x 720	6, 15, 30
848 x 480	6, 15, 30, 60, 90
640 x 480	6, 15, 30, 60, 90
640 x 360	6, 15, 30, 60, 90
480 x 270	6, 15, 30, 60, 90
424 x 240	6, 15, 30, 60, 90

💡 7.3 Three-Camera Bandwidth Limit and Recommended Settings

• With standard IPC + standard docking station, stable upper limit for three cameras: 640x480 @ 30fps.
• Default recommended setting: 424x240 @ 30fps (lower bandwidth usage, more stable).
• If you need higher image quality, prioritize reducing frame rate or reducing the number of concurrent cameras.