Skip to content

5.8dev merge for 4.0.0 and fab release - #179

Merged
WouterJansen merged 78 commits into
5.8from
5.8dev
Nov 16, 2026
Merged

WouterJansen merged 78 commits into
5.8from
5.8dev

Conversation

@WouterJansen

Copy link
Copy Markdown
Contributor

No description provided.

BeiZi6 and others added 30 commits September 11, 2026 19:21
Co-authored-by: Cursor <cursoragent@cursor.com>
…jansen2022

Register spawnObject actors for instance segmentation and material stencil
added third party licenses
…y for other simmodes outside the multirotor. This still disables debug drawing of pointclouds.
A vehicle that drops below the level's KillZ, for example driven off the
edge of the map or pushed through the floor after a collision, is destroyed
by the engine. AirSim keeps using the pawn, so the next sim update or reset
crashes (PawnSimApi::updateKinematics from CarPawnSimApi::updateRenderedState).

CarPawn and SkidVehiclePawn now override FellOutOfWorld to keep the vehicle,
log a warning, and move it back to its start pose on the next tick, with its
velocities zeroed. The teleport happens in Tick because FellOutOfWorld is
called from the physics sync, which overrides a teleport made there.

Found and tested in HERCULES (a fork of Cosys-AirSim) with Unreal Engine
5.8.3: a car teleported 500 m to 11 km below the Blocks map is back at its
start pose within 0.3 s and the simulation and reset() keep working.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JmWqXFCY6vDdpRe4VHoaQX
…path

moveOnPath, behind takeoff, moveToPosition and moveToZ, returned false for
moves that finished: its loop also ends when the drone stops within
getDistanceAccuracy() of the end of the path, where moveToPathPosition
stops commanding it, but only an exact arrival marked the task complete.
It now succeeds when the drone ends within twice the distance accuracy of
the end of the path (allowing for the drone still settling), and still
fails on a timeout or a drone stuck short of its goal.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JmWqXFCY6vDdpRe4VHoaQX
Each refresh hid every annotation component again in every camera's
scene and lighting captures and the player's view with AddUnique, which
is quadratic, and never dropped destroyed components; UpdateAnnotation-
Components also searched its list for every component. With thousands of
annotated meshes a refresh took tens of milliseconds to seconds of game
thread time.

- Hiding uses a set (linear) and drops destroyed components.
- The annotator remembers the components added since the previous
  refresh, and cameras hide only those; refresh serial numbers make a
  camera that is new, or missed a refresh, hide all of them instead. The
  player's view is updated once per refresh rather than once per camera.
- UpdateAnnotationComponents looks components up in a set.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JmWqXFCY6vDdpRe4VHoaQX
UAnnotationComponent ticks every frame, but only skeletal meshes need
per-tick updates; a static mesh's annotation is complete once its scene
proxy exists. With thousands of annotated static meshes the ticks cost
milliseconds per frame. Static-mesh annotations now disable their tick
once their proxy has been created.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JmWqXFCY6vDdpRe4VHoaQX
The LiDAR ran a ParallelFor per horizontal angle with one ray per
channel, so a 16-channel LiDAR at 10 Hz issued hundreds of tiny batches
per second and the calling thread spent most of its time handing out and
waiting for work. The rays of consecutive angles now go in one batch,
shot before a completed sweep is published, so the points are the same.

- The range noise generator was shared by the parallel rays (a data
  race); noise samples are drawn before each batch.
- The debug-point index used the 1-based step and wrote one row past
  the end of point_cloud_draw_ when DrawDebugPoints was on.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JmWqXFCY6vDdpRe4VHoaQX
Keep cars and skid-steer vehicles that fall out of the world
Report success from moveOnPath when the drone reaches the end of the path
Keep instance segmentation refreshes fast with many annotated components
Stop ticking annotation components of static meshes
Cast each LiDAR tick's rays in one parallel batch
…wList

- Restructure plugin Source/ into Source/AirSim/{Public,Private,Lib} plus
  Source/ThirdParty/{Eigen,rpclib,MavLinkCom} with their own Build.cs files
- build.cmd/build.sh stage AirLib headers/lib and third-party deps into the
  new layout; .gitignore and Blocks update scripts adjusted accordingly
- Add Cosys-Lab copyright header to plugin and AirLib sources
- Add PlatformAllowList (Win64, Linux, Mac) to the AirSim module
- Move update_mavlibkcom.bat out of AirLib to scripts/update_mavlinkcom.bat

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The engine ships its own "Eigen" module, so a project-level module of the
same name shadowed it and UBT refused to build ("Module 'Chaos' (Engine)
should not reference module 'Eigen' (Project)"). Rename the ThirdParty
modules to AirSimEigen, AirSimRpclib and AirSimMavLinkCom (folders,
Build.cs, build scripts, .gitignore) so they cannot collide with the
engine or other AirSim-derived plugins.

Also move Shell32/dinput8/dxguid (Win64) and stdc++/supc++ (Linux) to
PublicSystemLibraries to remove the "not resolvable to a file" warnings.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… warnings

- .cpp files in Private/<Sub>/ included their own header by bare name, which
  only resolved while .cpp and .h shared a folder. Include them by their path
  under Public/ instead (e.g. "Vehicles/Car/CarPawn.h"); 58 includes in 46 files.
- "AirLib/include/common/AirSimSettings.hpp" -> "common/AirSimSettings.hpp":
  AirLib headers are now staged directly into Public/AirLib (5 Beacon headers).
- Add f suffix to float literals flagged by C4305 (GPULidarSimpleParams,
  LidarCamera, PIPCamera).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Add f suffix to 23 double literals assigned/passed to float parameters
  (sensors, beacons, vehicle pawns, SimModeBase, EchoSimpleParams).
- DirectInputJoyStick: swscanf -> swscanf_s (Windows-only code path; %lX
  takes no buffer-size argument, so behaviour is unchanged).

Fab requires plugins to build without consequential warnings.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ocks

- AirLib.lib was built with whole-program optimization, forcing UBT's link
  of UnrealEditor-AirSim.dll to restart with /LTCG ("module compiled with /GL
  found"). /GL objects are also tied to the exact MSVC version that produced
  them, which breaks linking for users of a prebuilt AirLib.lib (Fab) on a
  different toolset update. RelWithDebInfo already had it disabled.
- Blocks DefaultEngine.ini: remove MinimumiOSVersion=IOS_11, which UE 5.8
  rejects at startup (LoadConfig import failed). Blocks does not target iOS.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The cvar was set twice: Set(3) at SetByCode priority, then via the
"r.CustomDepth 3" console command (SetByConsole). From the second PIE
session on, the SetByCode call was rejected and logged "Setting the
console variable 'r.CustomDepth' with 'SetByCode' was ignored". Set it
once with ECVF_SetByConsole when it is not already 3, and null-check
the cvar.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The high-poly SUV assets are downloaded by build.cmd/build.sh and are not
tracked; ignore them at their location after the Fab content reorganisation.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
These assets depend on the UE4 PhysXVehicles plugin (TireConfig,
VehicleAnimInstance), cannot be loaded in UE5 and are not used by
anything (the live BoxCar/SUV pawns use Chaos with the C++ wheel classes):
- VehicleAdv/Vehicle/* ("BareboneCar": pawn, anim BP, mesh, skeleton,
  physics asset, collision, tire configs) and its 2 joystick textures
- BoxCar: BoxCarAnim, Back/FrontWheel, Back/FrontTireData
- AirSimAssets.umap (old cook-everything map importing PhysXVehicles);
  DirectoriesToAlwaysCook already covers cooking the plugin content
(The 5 equivalent SUV leftovers are removed from the untracked download.)

Also delete 3 unreferenced redirectors whose targets do not exist
(M_DmS_Pallet_Rack_001, M_DmS_WireShelf_001 -> another project's /Game;
CameraLensDistortion_Inst -> a deleted material), drop the BareboneCar
default pawn path (AirSimSettings.hpp, docs/settings.md) and the
AirSimAssets MapsToCook entries (Blocks DefaultGame.ini, docs).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…refixes

Content reorganisation in UE 5.8:
- All 1372 plugin assets now live in Content/CosysAirSim/<Type>/<Group>/
  (Blueprints, Widgets, Materials, MaterialFunctions,
  MaterialParameterCollections, Meshes, Animations, Physics, Textures,
  Data, Sounds, Particles, Fonts), renamed with Epic type prefixes
  (BP_, WBP_, ABP_, M_, MI_, MF_, MPC_, SM_, SK_, SKEL_, PHYS_, PM_, T_,
  SW_, SC_, ATT_, PS_, Font_, DA_, Curve_, TI_). Every asset was re-saved
  in 5.8 (fixes "saved with empty engine version" warnings).
- Every /AirSim/... asset path in C++, AirLib default pawn paths, Blocks
  config and docs rewritten from the applied mapping
  (1083 replacements, verified).
- Config/DefaultAirSim.ini: CoreRedirects (package, object and Blueprint
  class) for every old path, so existing settings.json PawnBP entries and
  user levels keep working.
- Blocks FlyingExampleMap re-saved with the new references.
- seg_color_palette.png and irPalette.png (source images, not assets)
  moved to docs/images; Content/.gitignore removed (no loose files in
  Content).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
build.cmd and setup.sh now download cosys_car_assets_v2.zip (release
"carassets"), which holds the reorganised SUV under
CosysAirSim/<Type>/Vehicles/SUV, and extract it into Content/.
- Presence check uses the SUV pawn asset (BP_SuvCarPawn.uasset) instead
  of the empty VehicleAdv/SUV/v1.2.0 marker folder (Fab: no empty folders).
- Stale SUV copies are removed first, including a pre-Fab
  Content/VehicleAdv/SUV, whose old paths would clash with CoreRedirects.
- Messages updated: without the SUV the DefaultCar pawn is unavailable
  (the PhysX-era fallback vehicle no longer exists).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
WouterJansen and others added 28 commits November 10, 2026 09:13
#176)

Fixes #176

The annotation/segmentation setters, getters and list functions of
WorldSimApi called into ASimModeBase directly on the RPC thread.
updateAnnotation() then looked up the cameras again with
FindAllActor (a TActorIterator) on that thread, which asserts with
IsInGameThread(). The other annotator and capture list updates were
also racing with the game thread.

All these entry points now run as one task on the game thread, the
redundant actor lookups are removed and the camera count checks use
the right arrays. Regex calls now update all matches in one task.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Fixes #124

simSpawnObject cast the registry asset to UBlueprint, which is
editor-only data and is null in a packaged build, and then
dereferenced its GeneratedClass. The blueprint class is now resolved
from the GeneratedClassPath asset tag, the _C class path or (in the
editor) the blueprint itself, and cooked generated classes are listed
in the asset map as well. Names with a _C suffix are accepted, and a
wrong asset type gives an error instead of a crash.

Docs, the Python and MATLAB docstrings and the create_objects example
explain blueprint spawning and cooking unreferenced blueprints.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Fixes #109
Refs #107

FObjectAnnotator::UpdatePaintRGBComponent, UpdatePaintTextureComponent
and DeleteComponent searched the annotation component in all components
of the attachment root actor with substring matching. For meshes of
child actors nothing was found or a duplicate component was created,
and similar object or layer names could update the wrong component.
The component is now looked up among the children of the mesh with its
exact <layer>_<component> name, and UpdateAnnotationComponents matches
the exact names as well.

The example scripts, the ROS1 ground truth map and the MATLAB
getInstanceSegmentationLUT indexed the colormap by list position
instead of by the segmentation ID of each object. The docs and the
simGetSegmentationColorMap docstring explain the lookup.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Fixes #129

initialize() read the LockStep setting into lock_step_enabled_, but
openAllConnections() then called resetState(), which set it to false,
so lockstep mode could never become active. resetState() now restores
the value from the connection settings, also after a reconnect.

The lockstep docs describe when lockstep becomes active and that
messages sent to PX4 must use the PX4 time, and the settings example
uses the right LockStep key.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
)

Fixes #126

createDragVertices used pi*d^2 as rotor disc area (4x the real
pi*d^2/4) and pi*d*h as rotor side area (pi x the real d*h). The real
areas are now used, multiplied by two new parameters,
rotor_disc_drag_factor (4) and rotor_side_drag_factor (pi), that stand
for the extra drag of a spinning rotor. The defaults and the
multiplication order give bit-identical drag factors, so the top
speed, terminal velocity and wind response do not change.

angular_drag_coefficient is marked as unused. The flight controller
docs describe the drag model.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…er (#132)

Fixes #132

The PD position controller subscribed to odom_local_ned and
home_geo_point, while airsim_node publishes <odom_frame_id> (default
odom_local) and origin_geo_point, so it never got odometry. Its
velocity commands also could not work, as the odometry is published in
the ROS convention (y and z flipped from NED) while the velocity
commands were passed to AirSim in NED.

airsim_node now converts the world and body frame velocity commands
from the ROS convention (y left/west, z up, counterclockwise yaw rate)
to NED, matching the odometry. This is a breaking change for code that
sent NED commands.

PD controller fixes: topic names (odometry topic from the odom_frame_id
parameter), services under /airsim_node/<vehicle>/ as documented,
service responses filled and early returns on errors, GPS goals
converted to the odometry frame, absolute yaw error for reaching the
goal, derivative term on the change of the error with kd_yaw, yaw rate
limit applied to angular.z in rad/s, default control period, and the
dynamic constraints launch arguments passed to the node. The
dynamic_constraints launch file declared the gimbal yaw limit three
times, the last two are the roll limits. rviz and the docs use the
right topics.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Fixes #127

The LiDAR and GPU LiDAR PointCloud2 messages of the ROS2 node, and of
the ROS1 Python node, now have a uint16 ring field with the laser
channel of each point, 0 being the lowest laser. This is the field
name, type and numbering of the Velodyne and Ouster drivers, which
packages such as LIO-SAM need. The existing fields are unchanged and
there is no API change.

Both lidars give one point per laser for each horizontal step, so the
ring follows from the point index and NumberOfChannels: the CPU lidar
starts with the highest laser, the GPU lidar with the lowest one.

The lidar, GPU lidar, ROS2 and MATLAB docs describe the point order,
and the Python and MATLAB examples show how to get the ring from the
API data. The GPU lidar docs note that intensity needs
GenerateIntensity.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Fixes #113

The SkidVehicle SimMode only accepted the CPHusky and Pioneer vehicle
types. The skid-steer pawn already gets its vehicle API from
CarApiFactory, which creates an ArduRoverApi for ArduRover, and turns
the steering and throttle input into left and right wheel speeds. So
ArduRover is now accepted as vehicle type, using the CPHusky pawn
unless PawnPath is set.

The docs explain that ArduPilot has to use steering and throttle
outputs, not its own skid steering outputs.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Refs #125

moveByVelocityBodyFrameAsync(vx, vy, vz, duration, YawMode(...))
passes the YawMode as drivetrain. The simulator rejects the call, but
join() on the returned future does not raise that error, so the
command seems to do nothing. The move* APIs with a drivetrain or
yaw_mode argument now raise a TypeError for a wrong type, and the
API docs show passing yaw_mode as keyword argument. All example
scripts already pass the right types.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
spawnObject only made the requested name unique against the names
ListMatchingActors builds from mesh components, so actors without
meshes (for example a spawned light blueprint) were not seen. Spawning
a second object with the same name then hit Unreal's default fatal
error for duplicate actor names. The spawn now requests the name, so
Unreal makes it unique when needed, and the actual actor name is
returned and stored in the scene object map.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- AirLib's vendored nlohmann/json and CTPL headers ship in a new
  Source/ThirdParty/AirSimVendor module instead of inside the AirSim
  module (Fab: third-party code under Source/ThirdParty). build.cmd and
  build.sh copy them there; AirLib/include stays the source of truth.
  The unused bitmap_image, prettyprint and optional headers are removed.
- .uplugin: EngineVersion 5.8.0, CreatedBy Cosys-Lab, no Installed key.
- MavLinkCom public headers get the Cosys-Lab copyright notice.
- SimModeBase.h includes Particles/ParticleSystem.h instead of the
  monolithic ParticleDefinitions.h, which gave a warning in BuildPlugin.
- Removed the loose T_Coastline2.PNG source image from Content.
- build.sh puts the Linux/Mac static libraries in a Linux or Mac
  subfolder so one plugin can hold all platforms; the Build.cs files
  look there first and fall back to the old location.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The old pre-4.0 Calibri font face paths redirected to FontFace_Calibri
and FontFace_CALIBRIB, which are now FF_Calibri and FF_CalibriBold.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
rsync -a fails with "chgrp ... Operation not permitted" when the
repository is on a drive that does not support Unix ownership (for
example an NTFS or exFAT mount). The staging copies only need the files,
links and times, so use rsync -rlt.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
rsync -a fails with "chgrp ... Operation not permitted" or "failed to
set times ... Operation not permitted" when the repository is on a
drive that does not support Unix ownership or timestamps (for example
an NTFS or exFAT mount, or a mount owned by another user). The staging
copies only need the files and links, so use rsync -rl.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
CompressImageArray is deprecated since UE 5.1 and only calls
ThumbnailCompressImageArray; call that directly. Same behaviour, no
deprecation warning (seen in the Linux BuildPlugin).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
scripts/ros2_branch/make_ros2_branch.sh builds (or updates, one commit
per run) a "ros2" branch from any source branch with git plumbing,
without touching the working tree. It holds the ROS2 packages, AirLib,
MavLinkCom, the CMake files they need, docs/ros2.md, the licenses, and a
README, setup.sh (downloads rpclib and Eigen) and .gitignore from
scripts/ros2_branch/template.

CommonSetup.cmake also finds the repository root by
AirLib/include/common/Common.hpp, as AirSim.sln is not in that branch.
The ROS2 docs and the package README describe the branch.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ed command

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
rosdep only installs what package.xml declares, and cv_bridge,
std_srvs, OpenCV, PCL and yaml-cpp were missing, so colcon build failed
on a fresh machine (find_package(cv_bridge) not found). package.xml now
lists every find_package of CMakeLists.txt, uses the libpcl-all-dev
rosdep key instead of pcl, and drops the ROS1 message_runtime and
message_generation entries, so rosdep no longer needs --skip-keys.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Step-by-step setup from installing the plugin from Fab to flying a drone
from Python in your own project (GameMode, Player Start, settings.json,
keys, first script, packaging), an overview of the included content,
next steps and an FAQ. Linked from the docs home page, the README and
the navigation.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@WouterJansen WouterJansen changed the title 5.8dev 5.8dev merge for 4.0.0 and fab release Nov 16, 2026
@WouterJansen
WouterJansen merged commit ce7399d into 5.8 Nov 16, 2026
0 of 3 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Development

Successfully merging this pull request may close these issues.

4 participants