Troubleshooting#
Use this page when hardware setup, camera access, image streaming, or compute-unit performance issues block the Unitree G1 workflow.
Important
Stop the robot before troubleshooting hardware, cables, network connections, or camera access. Follow manufacturer safety procedures first.
Hardware Setup Reference#
Use the Isaac hardware setup troubleshooting guide for the source troubleshooting reference:
RealSense Camera Not Found#
Symptom#
Use this section when no RealSense camera is detected and one of these conditions occurs:
rs-enumerate-devicesfails withoutsudo, but succeeds withsudo rs-enumerate-devices.ros2 launch realsense2_camera rs_launch.pyreportsfailed to set power stateorThe requested device ... is NOT found.realsense-viewerdoesn’t detect any RealSense camera.
Solution 1: Check RealSense Versions#
Follow the RealSense setup instructions and confirm the installed versions for librealsense, realsense2_camera, and camera firmware.
Run the RealSense ROS node inside the Isaac ROS environment:
ros2 launch realsense2_camera rs_launch.py
Verify that the output reports the expected RealSense ROS and LibRealSense versions.
Check the camera firmware version:
rs-enumerate-devices
To update camera firmware:
Download the recommended D400 firmware binary from the Intel RealSense firmware releases.
Extract the binary from the downloaded
.zipfile.Run the firmware update command, replacing
<binary_filename>with the downloaded binary name:rs-fw-update -f <binary_filename>
Solution 2: Install RealSense udev Rules#
Verify that 99-realsense-libusb.rules exists in /etc/udev/rules.d/.
If it doesn’t exist, run these commands on the host machine before activating the Isaac ROS environment. Continue only after each command succeeds:
wget https://raw.githubusercontent.com/IntelRealSense/librealsense/v2.56.3/config/99-realsense-libusb.rules
sudo mv 99-realsense-libusb.rules /etc/udev/rules.d/
sudo udevadm control --reload-rules
sudo udevadm trigger
Confirm that each command exits without an error.
NVIDIA Isaac Capture UI Partially Disappears After Certificate Acceptance#
Symptom#
After accepting the CloudXR certificate in the headset browser, part of the NVIDIA Isaac Capture web client UI disappears or fails to render.
Solution#
Update the headset OS and native browser, then reopen the NVIDIA Isaac Capture web client and repeat the certificate acceptance step.
For PICO headsets: software version 5.15.5 was tested with this workflow.
MuJoCo Teleop Interferes With Real Robot Teleop#
Symptom#
When starting real robot teleop, a new MuJoCo window opens instead and the simulated robot falls because there is no virtual gantry.
Cause#
The MuJoCo teleop terminal or a background MuJoCo process is still running.
Solution#
Close the MuJoCo teleop window and stop the MuJoCo teleop terminal before starting real robot teleop. If the window is closed but the issue continues, stop any background MuJoCo processes and launch real robot teleop again.
Input Images Must Have an Even Height and Width#
Symptom#
Isaac ROS nodes terminate when given images with odd width or height. Terminal logs can include:
[NitrosImage]: [convert_to_custom] Image width/height must be even for creation of gxf::VideoBuffer
terminate called after throwing an instance of 'std::runtime_error'
what(): [convert_to_custom] Odd Image width or height.
Solution#
Replace the input image source with one that produces images with even width and height.
RealSense Does Not Stream IR Stereo Images#
Symptom#
Inside the Isaac ROS environment, realsense-viewer doesn’t show IR images, but depth images stream. Outside the Isaac ROS environment, no metadata containing projector status is attached to the IR frames.
Check the Supported Camera Stack#
Jetson Linux R38.4 uses Linux 6.8, as specified in its release notes. The Kernel 5.15 DKMS repair does not support this Thor kernel. No replacement kernel-module repair is qualified by this workflow.
Record the host kernel version with
uname -rand camera firmware withrs-enumerate-devices.Check the camera libraries and container layers against the Isaac ROS 4.5 RealSense setup and the workflow setup.
If IR streaming still fails, provide those versions and the camera logs to the hardware support owner before changing kernel modules. Check RGB streaming separately; an IR failure alone does not establish an RGB recording failure.
RealSense Error: Failed to Resolve the Request#
Symptom#
Any RealSense tutorial launches, but no images stream. Terminal logs include Failed to resolve the request.
Solution#
Confirm that the camera uses the correct firmware. Use the RealSense version and firmware checks in RealSense Camera Not Found.
RealSense Incompatible QoS Policy Error#
Symptom#
A RealSense output topic isn’t subscribed when a RealSense tutorial launches. Terminal logs report an incompatible QoS policy, such as RELIABILITY_QOS_POLICY.
Solution#
Set the QoS policy in the RealSense configuration file used by the launch graph so the RealSense publisher is compatible with the subscribing Isaac ROS node.
For the relevant RealSense image QoS setting, such as depth_qos or color_qos, use SYSTEM_DEFAULT.
Intel RealSense Camera Accidentally Enables Laser Emitter#
Symptom#
The Intel RealSense laser emitter turns on even though emitter_enabled is set to 0.
Solution#
Run this command at runtime:
ros2 param set /camera/camera depth_module.emitter_enabled 0
ROS_DOMAIN_ID Crosstalk#
Symptom#
Use this section when multiple robots, Thor systems, or ROS 2 groups are on the same network and you see unexpected ROS traffic:
ros2 topic listshows topics from another robot or workstation.Nodes connect to the wrong robot, camera, recorder, or deployment stack.
Teleop, recording, or deployment behavior is unexpected or changes when another ROS 2 system starts.
Cause#
ROS 2 uses DDS discovery. When multicast is supported on the network, ROS 2 systems on the same ROS_DOMAIN_ID can discover each other across that network. This can cause crosstalk when multiple robots or ROS 2 workflows run in the same lab.
For background, see the ROS 2 Humble documentation for ROS_DOMAIN_ID.
Solution#
Check the active domain ID. The default domain ID is
0.echo "${ROS_DOMAIN_ID:-0}"
Assign a unique domain ID before launching nodes:
export ROS_DOMAIN_ID=12
Confirm that only the expected nodes and topics are visible:
ros2 node list
ros2 topic list
Restart any launch files that were started with the wrong domain ID.
Intel RealSense D455 Infra Camera Capped at 15 FPS#
Symptom#
The Intel RealSense D455 infra camera is capped at 15 fps even though it’s specified to run at 90 fps.
Solution#
The initial frame rate can be capped around 15 fps due to a known RealSense ROS issue. Enable auto exposure:
ros2 param set /camera/camera depth_module.enable_auto_exposure true
System Throttled Due to Over-Current#
Symptom#
A display attached to the compute unit reports System throttled due to over-current.
Solution#
This warning indicates hardware throttling due to over-current. It can reduce performance and cause jitter or non-deterministic behavior in the robotics pipeline.
Reduce system load, such as lowering image resolutions or frame rates.
Warning
Disabling system over-current limits can allow the system to overheat and affect silicon health. Use this only when measuring appropriate limits for the workload.
To disable over-current throttling for the current boot, run:
sudo su
In the root shell, load the power-monitoring function:
source <(awk '/^function config_hwmon *\(\)/ {flag=1} flag; /^}/ && flag {flag=0}' /etc/systemd/nvpower.sh)
Then apply the setting in that same root shell:
config_hwmon ina3221 curr4_crit 81900 VDD_GPU_SOC
This setting does not persist across reboots.
Check how many times the system has throttled due to over-current since the last boot:
cat "$(find /sys/devices -name oc3_event_cnt -print -quit)"
If the command returns a non-zero value, the system has throttled due to over-current. The value indicates the number of throttle events.
G1 Workflow Checks#
Before changing G1 workflow commands, verify these G1-specific basics:
Confirm the G1 and NVIDIA Thor are powered on and connected through the required USB and direct Ethernet cables.
Confirm cable routing doesn’t interfere with the robot’s motion envelope.
Confirm the RealSense camera connection is visible to the system.
Confirm the Thor and G1 network interfaces are on the expected network.
Confirm the robot is in a safe state before restarting teleoperation, recording, or deployment services.