I am proud to present a new entry in the Videos section.
We demonstrate the Euclidean Distance Transformation feature of GPU-Voxels which allows you to derive 3D Distance Fields from pointclouds with high performance. In the video a flying drone plans a path through a environment, sensed only with on-board sensors. From the data a distance map is built on the fly with 30 Hz. So for every Voxel a simple wavefront planner can query the closest obstacle (plus its distance) and find a path that keeps a safety margin from the walls and only chooses openings that are large enough for the drone:
Try the feature by yourself! The X-Mas Release of GPU-Voxels comes with a “distance-kinect” example.
Thanks to Christian Jülg for re-implementing the “Parallel Banding Algorithm” that derives the Distance Maps! Here is a link to Cao Thanh Tung‘s original work: http://www.comp.nus.edu.sg/~tants/pba.html
Happy Holidays everyone!
Just for Christmas I finished the latest Release of GPU-Voxels which brings an amazing new feature:
Realtime distance field computation on large pointclouds (thanks to Christian Jülg)!!
Have a look at the Distance-Kinect example and have fun to inflate your avatar pointcloud (:
I already used this feature to implement a reactive inverse kinematics function for highly articulated robots, so that the robot maximizes free space which reaching a goal. Also we used the distance field to navigate a flying drone through a maze while keeping clearance to all obstacles (Video).
Of course there have been a bunch of Bugfixes and some more new features, so find the changelog below.
Download as always at: https://github.com/fzi-forschungszentrum-informatik/gpu-voxels
All the best to you for 2017,
V 0.93 (2016-12-24): ==================== API breaking changes: - Multiplying a vector<int> with an int scalar now gives a vector<int> - Removed offset when colliding VoxelLists as implementation was faulty - Unified CTor of VoxelLists and VoxelMaps - Unified spelling of insertPointcloud ==> insertPointCloud Major changes: - Added DistanceField calculations (thanks to Christian Jülg) - New DistanceVoxelMap type - Re-implementation of the Parallel Banding Algorithm (Original by Cao Thanh Tung) - Added according example and test cases - Added BitMasked collision check and testcase for BitVoxels - Implemented a HeighMap loader to populate 3D worlds from Bitmap files - Added insertPointCloudIntoMap() to HighLevel-API - Added collision functions between BitVectorVoxelList and BitVectorVoxelMap - Improved Mutex-Locking (thanks to Christian Jülg) - No more busy waiting - Recursive locking now possible - Reworked Boost-Tests to function as Benchmarks (Thanks to Herbert Pietrzyk) - Improved BitVector code - Fixed error in Bit-negation - Added according testcase - Fixed defaultCollider for BitVoxels (counted eBVM_FREE as collision previously) Minor changes: - Added screendump function for VoxelLists - Fix in CMake to export CUDA dependency - Restructured optional dependency management - Improved the handling of warnings within Kernels (less printf spam) - Visualizer config files: Added option to set color for all BitVoxelMeanings at once - Consistently using RPY for geometry generation (instead of YPR formerly) - Improved BitVoxelLists operations - Thrust performance optimization - Added according testcase Other changes: - Now compiles with CLang (fixed ambigous names) - Improved docs - Improved tescases for VoxelLists - Fix for CUDA 7 on Ubuntu 16.04 (-D_FORCE_INLINES) - Added UR10 & Robotique FTS model (10% larger, voxelized at 4mm) - Added Pelican UAV model - Corrected copy paste errors in CMake IC_BUILDER defines - Removed own source version of TinyXML: Now using system lib - Updated icmaker build tools - Updated QT include dirs in icl_core - icl_performance_monitor got some new functions
And again: There is a new version on GitHub.
It now centralizes the math functions directly where you would expect them: Within the datatypes.
Of course this breaks existing code, but it is straight forward to fix. Just call the functions which previously existed in the gpu_voxels namespace on the datatype (e.g. .toRPY() on a Matrix3f).
The datatypes itself stay untouched.
And have a look at the wonderful new Readme which now comes in Markup ((:
I just pushed the 0.91 release of GPU-Voxels to GitHub.
It now includes an example on how to link against the lib via plain CMake.
(It also fixes a minor bug and allows you to visualize pointclouds as PrimitiveArrays).
Hope that helps!
Finally I could release a new version of GPU-Voxels at https://github.com/fzi-forschungszentrum-informatik/gpu-voxels!
The release introduces a bunch of new features and makes everything more stable.
Also it compiles on Ubuntu 16.04!
Two little things are on the negative side:
– You probably need to patch your GLM library (see this issue: https://github.com/g-truc/glm/issues/530 )
– The update breaks some API functions, but it should be easy to adapt in your code.
Here is the rather long changelog:
V 0.9 (2016-08-21): =================== API breaking changes: - GpuVoxels is now a singleton and has to be initialized - BitVoxelMeaning enum changed! 0 = Free, 1 = Occupied. More SV IDs. - Added as() operator to cast general maps into specific maps - Map-Offsets may now be negative, so Voxel-pointers changed datatype. - Shifted main API from general map type to specific implementations: Many functions can no longer be called on basic map types but only on specific maps. As not all map types offer all interfaces, this allowed us to remove unimplemented function stubs (thanks to Herbert Pietrzyk) - RPY rotation order changed to ROS standards: First rotated around roll, then pitch, then yaw
Major changes: - Unified map-locking for all map types to guarantee thread-safety - Added new Pointcloud class for single clouds (thanks to Herbert Pietrzyk) - Added Octree-API function: insertPointCloudWithFreespaceCalculation to trigger raycasting - Added option to interpret unknown cells of an octree as obstacles when checking collisions - Added tfHelper class to interact easily with ROS tf - New Math functions: - Added host function to invert matrices. Code thankfully copied from Maxim Singer - Function to convert Mat4 to Roll, Pitch, Yaw together with Boost tests - Vector3f now offers: apprx_equal, length, normalize, dot, cross - angleBetween two vectors, orientationMatrixDiff between two matrices - Matrix4f now offers: equality, approximate equality and subtract together with Boost tests
Minor changes: - Added visualizer config file and a python generator for random swept volume colors (this time for real) - New Boost testcases for Pointclouds and MetaPointClouds - Simplified sensor code for Raycasting in Octree - Restructured keyboard shortcuts in visualizer: - Added "command mode" to switch between data types so all Function keys can toggle maps of the selected kind. - Using ALT+digit to set decimal preposition of SweptVol IDs - Right-Click available for more datatypes in Visualizer to print voxel information - Fixed updates of subclouds in MetaPointClouds - Added sanity check in computeLinearLoad - Added Getter functions for GVL parameters - Added some general HTML pages to Doxygen docu (thanks to Darius Pietsch) - Unified probability type in all maps - Fixed memory leaks in MetaPointCloud - Added Kernel for GPU memory comparison - Unified geometric transformation kernels - Clarified singed and unsigned voxel indices (thanks to Christian Juelg) - GPU Voxels main class now checks for Compute Capability at init - Added PointCloud constructor to load file
Other changes: - Compiles with Ubuntu 16.04 - Added CMake macro to remove VTK defines - Required lib glm fix: https://github.com/g-truc/glm/issues/530 - Added enlarged UR5 model - Updated list of contributors
I just wanted to announce two new videos:
|The YouTube video shows an application scenario with a shared workspace where the motions of the robot (represented as temporally encoded Swept-Volumes) are monitored via two Kinect cameras. As soon as an obstacle is detected in the upcoming trajectory, the robot stops and chooses another (collision free) subtask to execute. GPU-Voxels acts as some kind of relay between ROS MoveIt and the robot controller, so that planned motions can be monitored online.|
|This video (third one on the linked multimedia page) shows our work on predictive collision detection: We analyze motion in RGB-D video input and project objects movements into the future.
This generates a Swept-Volume that can be intersected with the Swept-Volumes of planned robot motions to anticipate probable upcoming collisions.
Big thanks to Felix Mauch for his great work on this toppic!
For further info please read our paper or contact me.
A. Hermann, F. Mauch, K. Fischnaller, S. Klemm, A. Roennau, R. Dillmann,
We are also making progress on bringing GPU-Voxels to a flying drone scenario, mobile platforms (motion primitive planning in simulation is already up and running) and into real-time grasp planning. So stay tuned!
Both is fixed with this release (:
V 0.82 (2016-02-12): ==================== - Fixed severe Bug in the locking of Voxellists. Warning: This only concerns VoxelLists. - Added collision check between ProbabilisticVoxelMaps and VoxelLists. - Added transform functionality to MetaPointClouds and removed it from Robot classes. - Fixed Vec3ToMat4() function. - Implemented Deep-Copy operator for MetaPointClouds. - Implemented Pointcloud-Offsets for geometry entities of URDF models. - Warning: Only translation, NO rotation available here! - Added new struct to describe oriented boxes. - Added new file for helper functions to generate pointclouds from geometric primitives. - Fixed RPY rotation matrices. - Added Boost testcase for shifted collision checks. - Added script to generate Visualizer configuration with random colors for SweptVolume IDs. - Refined the models of the SVH with 0.4 mm resolution.
V 0.81 (2015-12-14): ==================== - Fixed severe Bug in MetaPointCloud::updatePointCloud() that lead to Segfaults. - Implemented new functions for VoxelLists: - Equality check (Mainly for Boost-Tests) - Subtraction (Bool operation) - Collisioncheck that counts the collisions per Voxel-meaning. - ShrinkToFit: Frees preallocated Thrust memory - Fixed error handling when reading and writing all data structures. - Added more Boost testcases. - Fixed URDF robot rendering when for collision meshes. - Made some const functions const. - Visualizer: - Improved python script to generate color gradients for Visualizer configfiles. - Made Voxellist Voxels clickable - Enhanced reliability of Visualizer (Thanks to Christian Juelg) - Added slicing functionality to Visualizer (Thanks to Christian Juelg) - Improved voxelize.sh script to delete intermediate files.
Next round: Release 0.8 is available!
After a longer time without any news, here is a pretty long list of updates! We are now compatible with CUDA 7.5 (and will officially only support that from now on). On the content side, we implemented Voxellists and extended the Bitvoxel-Collision-Checks. As the lists only safe occupied Voxels, they give a huge speed-up when dealing with sparse data. So especially the combination of representing a Robot and its plans in Voxellists and look them up in the Octree of the live environment is a good practice!
Thanks to Felix Mauch it is now possible to have Bitvoxel-Collision checks that regard a “Bit-Margin”. So if the bits represent a point in time, this way a “fuzzy” collision check can be realized.
We have also done a lot of restructuring and optimization work to improve the overall software quality, fixed a bunch of bugs, and dropped our own copy of the CUB library in favour of Thrusts copy of the lib, that gets installed with CUDA.
Within FZI we have a growing list of use-cases, demos and extensions that heavily rely on GPU-Voxels. So I hope to publish some news on that soon, including motion planning with motion primitives, grasp planning and scheduling tasks in multi-robot scenarios.
Hope you like it and that it will help in improving your robots!
V 0.8 (2015-11-09): =================== - Compiles with CUDA 7.5 - Added Voxellists as new Datatype - Huge speedup compared to Voxelmaps with sparse data. - Added Bit-Shifting-Collisionchecks (Thanks to Felix Mauch) - Added Bit-Margin for Bitvoxel Collisionchecks (Thanks to Felix Mauch) - General: - Removed superficial memory allocation from Voxelmap. - Added script to automatize binvox mesh voxelization - Improved PrimitiveArrays update speed when number of primitives stays constant. - Removed heaps of compiler warnings (Thanks to Christian Juelg) - Various Bugfixes, mainly in collision Kernels - Renamings / moved code: - HighLevel-API: Added "merge" function. - CudaMath now longer requires an instance. Now it is a namespace. - VoxelType ==> BitVoxelMeaning - Fixed standard CTORs of basic datatypes for CUDA shared memory allocation. - Moved voxels to gpu_voxels namespace - Removed seom missleading typedefs - Moved Voxel definitions to own subfolder. - Removed dead code fragments. - Octree: - Adapted to new API. - Removed its own voxellist. Now uses more general Voxellist. - Removed VoxelTypeFlags and replaced them by more general bitvectors. - Simplified Octree selftests and triggered them by Boost-Test. - Dropped CUB lib. Now using CUDAs version from Thrust. - Robot representation: - Enhanced robot interface (jointnames and min max angles). - Restructured inheritance of robot implementations. - Fixed install of robot headers. - Visualizer: - Updated Visualizer to new API. - Added new function to move cam in plane. - Switched from Octree to more general Cubelist interface, as this can also be used to visualize Voxellists. - Visualizer now cleans Shared-Mem files. - Added Script to generate color gradient config files for visualizer. - Fixed a bug in drawing the grid - Example programs and Testcases: - Updated examples to new API. - Added more Boost-Tests. - Reordered model files. - icl_core and basic libs: - Added performance monitor docu - Some new cmake magic - Improved ringbuffer - New Doxygen version. - Tinyxml cmake updated. - Added missing Findscripts for OpenNi.
We have published a new work on predictive collision detection using GPU-Voxels. The paper should be online within this year at IEEE-Explore and on my Research-Gate profile.
The contribution is on the combination of RGB-D Flow to create Swept-Volumes out of predicted motions in a pointcloud and check those for collisions with Swept-Volumes of planned robot motions. That way we can trigger replanning steps around a moving obstacle or stop the robot, before collisions can occur.
A. Hermann, F. Mauch, K. Fischnaller, S. Klemm, A. Roennau, R. Dillmann,
“Anticipate your surroundings: Predictive collision detection between dynamic obstacles and planned robot trajectories on the GPU”
in European Conference on Mobile Robotics (ECMR 2015, Lincoln, September 2015)