Frequently Asked Questions

This section provides answers to frequently asked questions about NVIDIA Isaac SDK™. Use it as the first step in troubleshooting problems.

To ask and answer questions in the Isaac SDK developer community, visit the Isaac SDK Forum.

Hardware Components

When I run my application that uses a ZED Camera, I get the following error:

engine/alice/backend/modules.cpp@74: dlopen failed: libhidapi-libusb.so.0: cannot open shared
object file: No such file or directory

The libusb-dev library is not installed. Install libusb-dev and all other required dependencies on the robot with the following command:

bob@desktop:~/isaac$ ./engine/build/scripts/install_dependencies_jetson.sh -h <jetson_ip> -u <jetson_username>

Where <jetson_ip> is the IP address of the robot.

I am running my app on Kaya. Everything runs fine, but the joystick does not move the robot.

Make sure you press the “deadman switch” L1 on the joystick while using the direction knobs. Pressing L1 is required to prevent unwanted movements.

When I run my application that uses a BMI160 IMU, I get the following error:

I2C Error: Device or resource busy (errno 16)

The kernel module nvs_bmi160 is using the I2C resource needed by the application. Unload the module with the command sudo rmmod nvs_bmi160. You can also run the install_dependencies_jetson.sh script, which prevents nvs_bmi160 from loading.

bob@desktop:~/isaac$ ./engine/build/scripts/install_dependencies_jetson.sh -h <jetson_ip> -u <jetson_username>

Visualization and WebSight

I don’t see images in Sight. Or the images are blank. Or the labels displayed are red/yellow.

There are a number of reasons why a channel may not update correctly:

  • The channel is not enabled in the bottom left menu. If the channel is not enabled, the robot does not send the update to the frontend and the channel appears in red in the renderer’s legend. Enable this channel from the channel menu directly or right click on the renderer and click enable all channels.

  • The channel is enabled, but there is no valid transformation between the channel and the reference frame used by the renderer. If this is the case, the channel appears in yellow in the renderer’s legend. If you hover over the channel name, additional information is provided about the missing transformation: ‘[reference_T_channel] is not defined’. You can check the PoseTree widget to figure out if there is a path between both coordinate frames.

    Note

    ‘’#####’’ is used as a default frame. It means that no frame has been provided for the channel.)

  • Your image/channel is rendered, but a later channel is overriding it. Channels are rendered in the order they appear in the legend: the top one first, followed by second one, and so on. If you render an image the size of the renderer, it covers all the previous channels. Make sure your channels are in the right order. You can change the order by right clicking the renderer and clicking settings. From there you can use the arrows on the left side of the channels to update the order.

  • The timestamp of the channel is in the future. Each show operation in Sight has its own timestamp (if none is provided, the current time is used). In general, it is good practice to use the acquisition time to render images or objects; it helps to synchronize the channels. Sight uses the application time (seconds since the beginning) to render the different channels. If the timestamp of a channel is in the future (either because the acquisition time used is not relative to the start of the application or because the wrong unit is used), then the channel is not updated until the application time catches up with the channel.

    This case can be detected by refreshing the web page. If the image/channel updates but then freezes, it means that Sight received the latest message but is waiting for the time to catch up before rendering.

    If you own the codelet rendering the image, make sure that you are using the right time in your show call. If you do not own the codelet, then the acquisition time of one of the channels is most likely used to render. Make sure the publisher provides the right acquisition time.

I am having lag issues in Sight.

See below for different types of lag issue:

  • All the Sight widgets are lagging. You are most likely streaming too much data to Sight and your network connection can’t handle the load. First, try enabling only the channels you are interested in visualizing. To achieve this quickly, first unselect all the channels by clicking on the root of the channel list (left panel). You can then re-enable all the channels of a given renderer by right clicking on it and selecting Enable all channels.

  • I enabled only the channels I need, but Sight is still lagging. You are probably looking at a channel that requires a lot information. It could come from either a channel with a lot of drawing primitives (PointCloudViewer for example) or large images. Most of the Isaac default viewers have an option to downscale the amount of data sent to Sight (CameraViewer has a reduce_scale parameter and PointCloudViewer has a`skip` parameter).

  • Sight is rendering with a decent framerate, but some channels seem to be lagging behind. By default, each renderer renders with a delay of 0.2s to allow all the channels to arrive. However, 0.2s might not be enough for channels where the duration from sensor acquisition to Sight is more than 0.2s. You can adjust the rendering delay of each renderer individually by right-clicking the renderer and hovering over the Change delay option: a slider will appear, allowing you to adjust the delay from 0 to 2s.

  • The bounding box for ML detections do not match the image. If you have tried the solutions above and they did not help, check that each detection matches an image and each image has its own detection. If you reduced the frame rate of the CameraViewer, it is possible a detection is being rendered without its matching image. On the other hand, if the detection pipeline can’t keep up with the camera framerate, it is possible some frames are not being processed. Try reducing the frame rate of the camera to make sure no frames are skipped.

Unable to see the Sight webpage after opening http://localhost:3000 in a browser.

Make sure an application is running on your desktop. If you are running the application on the robot, you must use the robot IP address.

If the “Failed to start Webserver!” error message appears in the logs, the port may not have been released by a previous application.

Use the following command to determine the application using the port:

lsof -i TCP:3000

You can then kill the application with the following command:

killall -9 <app name>

If you need to keep the app running, update the configuration file to change the port to another open port.

Sight visualization is very slow.

When network bandwidth is insufficient, channels displayed in Sight may exhibit latency or very low framerate. Follow these steps to improve bandwidth:

  1. In Sight, under Channels, uncheck all unnecessary channels.

    Tip

    You can disable all channels at once by clicking on the application name in the channel menu. Enable all channels you want to visualize by right clicking a renderer and selecting enable all channels.

  2. Verify Wi-Fi antennas and cables are connected to the PCIe Wi-Fi card.

  3. Determine whether power-saving mode is enabled on the Wi-Fi adapter:

    cat /etc/NetworkManager/conf.d/default-wifi-powersave-on.conf
    [connection]
    wifi.powersave = 3
    

    Set the wifi.powersave value to 2 to disable power-saving mode:

    wifi.powersave = 2
    
  4. Try a wired connection to rule out Wi-Fi bandwidth issues.

  5. If Sight has been running for a long time, it might have accumulated a lot of data, which slows down the frontend. Try refreshing from time to time to see if it improves performance.

Other Questions

What is the easiest way to generate a capnp id for a new message?

Every capnp file requires a unique ID at the beginning of the file. If you create a new capnp file and attempt to build Isaac without an ID in your new file, an ID is generated and printed in the error message, similar to the following:

messages/my_new.capnp:1:1: error: File does not declare an ID. I've generated one for you.
Add this line to your file: @0xcdeac1e381086f01;

As instructed by the error message above, add an “@” symbol, then the generated ID, then a semicolon to the top of your capnp file (if you don’t add the semicolon, you will recieve a parse error).

When running an application that requires TensorFlow, I get the following error: PANIC engine/core/buffers/algorithm_cuda.cpp@55: Could not copy memory. Error: 35

To resolve this error, install CUDA 10.0 using the instructions at NVIDIA CUDA Installation Guide for Linux.

When using a USB Camera on a TX2 I see no image or a heavily distorted image.

This happens when the CPU clock speed is too low on the TX2 and it cannot process the USB frames in time. Try running the jetson_clocks command to increase the CPU clock speed. See this NVIDIA Developer Forums topic for more details.

What is the syntax for setting a Pose3d in json?

Earlier versions of Isaac SDK expected an array of 7 numbers to set a Pose3d in JSON. The old syntax, which is still supported, looks as follows:

[qw, qx, qy, qz, x, y, z]

The first four numbers form the quaternion while the latter three are the translation values.

Some users may find the new format more convenient. For details on the syntax, please see engine/gems/serialization/json_formatter.hpp and for examples please check engine/gems/serialization/tests/json_formatter.cpp. Below are some equivalent poses:

old syntax (supported) Use square brackets

new syntax Use curly brackets

[0.0, 0.0, 0.0, 1.0, 25.0, 20.0, 0.0]
{
   "translation": [25.0, 20.0, 0.0],
   "rotation": {
      "qw": 0.0,
      "qx": 0.0,
      "qy": 0.0,
      "qz": 1.0
   }
}
[0.0, 0.0, 0.0, 1.0, 25.0, 20.0, 0.0]
{
   "translation": [25.0, 20.0, 0.0],

   "rotation": { "yaw_degrees": 180 }
}
[0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0]
{
   "rotation": { "yaw_radians": 3.14159 }
}
[1.0, 0.0, 0.0, 0.0, 25.0, 20.0, 0.0]
{
   "translation": [25.0, 20.0, 0.0]
}

Troubleshooting

What should I check first if my build is failing?

Make sure you are running Ubuntu 18.04 on your host system, and that you have run the install_dependencies.sh script.

Why am I seeing “crosstool” errors when running Bazel build?

If you see the following errors during Bazel build, your Bazel version is out of date:

ERROR: @bazel_tools//tools/build_defs/cc:action_names.bzl' does not contain symbol 'ACTION_NAMES
ERROR: error loading package '@toolchain//crosstool': Extension file 'crosstool/cc_toolchain_config.bzl' has errors

Use the engine/build/scripts/install_dependencies.sh script to upgrade Bazel to the current version required by Isaac SDK. To perform a manual update, see the GitHub page for the Bazel installer.

When using a USB Camera on a TX2 I see no image or a heavily distorted image.

This happens when the CPU clock speed is too low on the TX2 and it cannot process the USB frames in time. Try running the jetson_clocks command to increase the CPU clock speed. See this NVIDIA Developer Forums topic for more details.

I am seeing an error similar to “libnppc.so.10.0: cannot open shared object file: No such file or directory”

Some Isaac SDK code requires CUDA 10.0 to execute properly. Ensure CUDA 10.0 is installed locally and the /usr/local/cuda-10.0/lib64 value is part of the LD_LIBRARY_PATH environment variable.

I am seeing “Could not deserialize configuration parameter” errors when I run an application.

Configuration values must match the data type specified in the component API. See the Component API Overview or the component .hpp file for the expected data type.

Note that an integer value is accepted as a type of double value.

How to Add a New External Dependency

Your packages or components may require external dependencies not yet available in Isaac. The right strategy for integrating a new external dependency can vary case by case. There are several examples in the third_party folder. This tutorial explains how to add the zlib library.

  1. Find a reliable source for your dependency. In the case of zlib, get it from the official zlib webpage at zlib.net. To avoid unexpected problems in your codebase when the external dependency is updated, fix a specific version of the library. This tutorial chooses the version 1.2.11.

  2. Add the dependency to your WORKSPACE file using the Bazel rule isaac_http_archive:

    isaac_http_archive(
        name = "zlib",
        build_file = clean_dep("//zlib.BUILD"),
        sha256 = "c3e5e9...cb1a1",
        strip_prefix = "zlib-1.2.11",
        url = "https://zlib.net/zlib-1.2.11.tar.gz",
    )
    

    Importing external dependencies as an archive is generally preferred over using the rules new_git_repository or git_repository, which import git repositories, because archives are smaller and have less overhead than git repositories.

  3. For the purposes of this tutorial, assume that you have to write a BUILD file for the new external dependency (this may not always be required). In the case of zlib, this is quite straightforward:

    cc_library(
      name = "zlib",
      srcs = glob(["*.c", "*.h"], exclude="zlib.h"),
      hdrs = ["zlib.h"],
      copts = [
        "-Wno-shift-negative-value",
        "-Wno-implicit-function-declaration",
      ],
      includes = ["."],
    )
    

    This BUILD file defines a single C++ library named zlib using the Bazel rule cc_library. The library includes all C source and header files from zlib and compiles them into a single library. The only external header file is zlib.h, which is excluded from the sources so that the same file does not appear as a source and a header.

    In general, srcs is used for files internal to the library and can be used for both source and header files, while hdrs is used for publicly facing header files that are required in applications that use the library.

    A couple of compiler flags are also required. Add them using the copts option to avoid compiler warnings that are treated as errors due to use of the -Wall option by the Isaac compilation tool.

The new library is ready for use as a dependency in one of your libraries or binaries, similar to the following example:

cc_library(
  name = "foo",
  srcs = ["foo.cpp"]
  hdrs = ["foo.hpp"],
  deps = ["@zlib"]
)

Every external dependency creates its namespace with the same name as the one used for the isaac_http_archive or corresponding rule. In this case, the name of both the dependency and the library is zlib and thus the shortcut @zlib can be used to refer to the library. If there were a second library foo inside the same archive, the explicit form @zlib//:foo would be needed to reference it.

How To Analyze a Crash using Minidump

In case of a crash, Isaac SDK applications attempt to collect information with breakpad. This information is written into a minidump file in binary format. This section explains how to extract human-readable information from minidump files.

Prepare the Minidump Toolkit

Prepare the minidump toolkit in /tmp/minidump/ with the following command:

bob@desktop:~/isaac$ ./engine/build/scripts/prepare_minidump_tools.sh

Prepare Symbols

Prepare symbols before testing applications by passing the argument -s when deploying applications as shown in the following command:

bob@desktop:~/isaac$ ./engine/build/deploy.sh -s -h <robot_ip> -d <device> -p <target>
--remote_user <username_on_robot>

where <robot_ip> is the IP address of the robot and <username_on_robot> is your username on the robot.

Analyzing Minidumps

When a crash occurs, Isaac application minidump file paths are reported on the console similar to the following:

Minidump written to: /tmp/28db00a1-e756-4c47-62f6a7b6-fc26c1a0.dmp

To make the minidump data human-readable, run the following command on a desktop that has symbols prepared:

bob@desktop:~/isaac$ ./engine/build/scripts/process_jetson_minidump.sh -h <robot_ip> -d
/tmp/28db00a1-e756-4c47-62f6a7b6-fc26c1a0.dmp --remote_user <username_on_robot>

where <robot_ip> is the IP address of the robot and <username_on_robot> is your user name on the robot. If the minidump file is locally available, run the following command:

bob@desktop:~/isaac$ ./engine/build/scripts/process_minidump.sh <minidump>

Crash information is presented on the console, and is similar to the following:

Crash reason:  SIGSEGV /SEGV_MAPERR
Crash address: 0x0
Process uptime: not available

Thread 11 (crashed)
 0  realsense_camera!isaac::RealsenseCamera::tick() [RealsenseCamera.cpp : 533 + 0x4]
     x0 = 0x0000000000000000    x1 = 0x0000007f962b2180
     x2 = 0x0000007fa0000080    x3 = 0x0000007f800008d0
     x4 = 0x0000000000000007    x5 = 0x00000000007e8ac8
     x6 = 0x00000000007e89f0    x7 = 0x0000000000063762
     x8 = 0x00000000000000d7    x9 = 0x001dcd6500000000
    x10 = 0x000000005c787d6b   x11 = 0x000000000f171c90
    x12 = 0x0000000000000017   x13 = 0x000000005c787d6b
    x14 = 0x00076f61bec52c49   x15 = 0x00001c2c8948c9df
    x16 = 0x0000007fa7708688   x17 = 0x0000007fa7471838
    x18 = 0x0000007fa005bdcc   x19 = 0x0000000037cc4610
    x20 = 0x0000007fa004bbb0   x21 = 0x0000007fa0248c01
    x22 = 0x0000000037cc47e0   x23 = 0x00000006556cdce8
    x24 = 0x000000000091f000   x25 = 0x0000007fa53dd330
    x26 = 0x0000000000000000   x27 = 0x0000007f962b31b0
    x28 = 0x0000000000000001    fp = 0x0000007f962b20a0
     lr = 0x0000000000459f6c    sp = 0x0000007f962b20a0
     pc = 0x0000000000459f74