Difference between revisions of "User:Abhijit"
|Line 38:||Line 38:|
#Finishing the logic for creating manifolds using rays shot in the direction of the positive X axis.
#Finishing the logic for creating manifolds using rays shot in the direction of the positive X axis.
Revision as of 13:18, 21 October 2011
- 1 ESA Summer of Code Project Proposal
ESA Summer of Code Project Proposal
Non-vacuum gravity simulator for the BRL-CAD Solid Modeling tool
- Name: Abhijit Nandy
- Mailing List ID: firstname.lastname@example.org
- IRC ID: abhi2011
I am an international student (nationality Indian) in the second year of the Computer Engineering programme at TU Delft, The Netherlands. I have a computer science background and was working in the software industry for three years from 2006 to 2009. After that I decided to pursue a course in Computer Engineering to get acquainted with hardware development. More details are present in my CV. Currently I have finished my courses and thesis work and so I have free time on my hand to devote 40 hours a week to the development of BRL-CAD.
Brief project summary
Currently BRL-CAD provides facilities for representing geometry, but has limited capability for simulating physics effects on that geometry. The facilities currently present were designed with animations in mind. The proposal is to integrate rigid body dynamics using the Bullet Physics Engine or the Open Dynamics Engine (ODE). Bullet is provided under the zlib license while ODE is provided under both the BSD license and the LGPL. Both engines provide a collision detection engine and a Newtonian physics model. The exact physics engine to be used will be researched in the first week (see Development timeline). The chosen engine’s collision detection mechanism will be used to check for overlap of axis aligned bounding boxes or the bounding sphere. If an overlap is detected more accurate collision detection will be delegated to a custom collision handler that may use the BRL-CAD gqa and rtcheck tools for precise detection. Other collision handlers can be added if required in case they are absent for any of the BRL-CAD primitives. After the collision handling is tested to work correctly with the rest of the system, support will be added for user specified forces apart from ambient gravity. If time permits I intend to work on integrating and improving the system for joining objects together as well(see Detailed project description).
Detailed project description
The physics will be implemented as a command in the BRL-CAD Archer GUI. This will be of the form:
Archer> runphysics N
The parameter N is the number of simulation steps to run the physics. After the user invokes this command, the Tcl code for the command will call a C++ function containing the logic for doing the physics simulation. The function can be of the form:
int ged_runphysics(struct ged *gedp, int argc, const char *argv)
The following occurs in the C++ function which will be central to the implementation:
- The struct ged object that is typically passed to plugins will be used to build up the physics world in a btDiscreteDynamicsWorld provided by Bullet.
- A Bullet collision shape will be inserted into the physics world corresponding to each object present in the struct ged. These are based on convex primitives such as btBoxShape, btSphereShape, btCylinderShape and they can approximate the BRL-CAD objects. The collision shapes will be constructed from BRL-CAD geometry information.
- After this the force vectors requested by the user will be read either from the struct ged or using other LIBGED commands. These will be applied on the collision shapes in the physics world.
- The physics world will then be stepped just once and the ged object positions will be set to the new positions using information from the Bullet btDiscreteDynamicsWorld.
- The repositioning of objects in Archer happens automatically by the command wrapper, once the objects are modified using LIBGED commands.
- The C++ function will then exit and control will return to Tcl.
The Tcl command wrapper will call the C++ function N times to obtain the simulation result after N steps. As the object positions will be updated on each call to ged_runphysics(), the user will be able to see the scene’s objects animate in Archer.
Particularly with regard to Bullet, as that is the only physics engine I have used so far, I am aware that it provides facilities for contact tests  and callbacks using the contactTest and contactPairTest queries. These can be accelerated using faster broadphase algorithms such as btDbvtBroadphase or btAxisSweep3 which affects the speed of test for overlapping AABBs . Therefore real time updates should be possible. Moreover there are facilities for more fine grained control over the collision pipeline of Bullet, utilizing a custom class derived from btCollisionDispatcher . Custom collision detection algorithms can be specified using btDispatcher::registerCollisionAlgorithm .
Physics engines also typically provide constraints such as point to point, hinge and 6 degree of freedom constraints which can be used to join objects together and improve the rudimentary joint rigging systems currently present.
Note: The implementation will not rely on any polygonal mesh data as BRL-CAD does not store a mesh for its primitives. Instead physics objects for Bullet’s dynamics world will be created from other geometry information e.g. radius and tire width can be used to make a bounding cylinder for a tire.
- Oct 20th
- Adapting the collision logic to use forces instead of contact pairs. Simply #adding contact points between objects does not seem to cause Bullet to apply reaction forces between bodies
- So now forces are applied through the center of the contact region and normal to the surface of either body, which should be enough to push them apart.
- Friction forces need to be applied parallel to the surface too, which needs to inserted ,otherwise we won't have balls rolling on planes.
- Oct 16th
- Added contact processing callbacks to investigate the issue with forces not being applied after contact pairs between 2 objects are added to the physics world.
- Got rid of multiple linked lists which were put in to support varying lists of manifolds, but which turned out to be more of trouble than its worth
- Oct 12th
- Finishing the logic for creating manifolds using rays shot in the direction of the positive X axis.
- Oct 7th
- Grinding out memory leaks using valgrind !
- Lesson : when in doubt always free memory and cause a double free crash, better to be sure!
- Raytracer is properly recording overlap areas and hit regions now.
- Oct 6th
- Hmm things getting even more interesting, it seems that there can be extremely thin overlap regions between objects which are resting on top of each other, down to 0.000001 m, so a plane with all the manifold points can be generated for such a case, as a cuboidal regions would be too small for the solvers to analyze accurately(need to verify this).
- Added code to actually shoot the rays now and record the overlap regions in a circularly linked lists, needs some debugging however.
- Oct 5th
- Began writing a manifold generator based on raytracing. The rays will be shot only where broadphase collision detection finds AABBs overlaps
- This will allow higher accuracy with rays being shot in a 3D grid of higher granularity without sacrificing performance.
- Oct 4th
- Added code to display manifold normals correctly.
- Figured out how to generate the normals based on the direction in which objects overlap. This will be required when generating the contact manifolds using raytracing.
- Oct 3rd
- Contacts manifold are getting drawn and will help in reproducing them through raytracing.
- Corrected some code to position objects correctly and stopped creating bounding boxes every iteration as it has been causing issues.
- Sept 30th
- Added code to display bounding boxes, recalculating the BB every iteration is causing boxes that have moved to stay bent in their new position as each ounding box is inserted as a new cube in the next iteration. This problem should disappear when contact points are calculated correctly.
- Sept 25th
- Finished reorganizing the code to integrate raytracing. Ran into some issues with the positioning of objects after each physics steps, so had to revert to some older working copies.
- Basically at the beginning of each step I rebuild the world and after the simulation, I save the world transform and velocities(linear and angular) for initializing the next step.
- Each step has the bounding box for all objects in the scene calculated again as was originally planned.
- Sept 16th
- On the other hand it seems better to add a new collision detection algorithm for box-box collisions.
- Bullet allows custom algos for each combination of colliding shapes. However we are interested only in the box-box case since all arbitrary shapes are represented by the btBoxShape collision primitive in Bullet. However manipulation of the contact points according to rt output should still allow correct behaviour of arbitrary shapes.
- Inserted some framework code for a basic rt based collision algorithm.
- The contact manifolds will be generated by rt, the algorithm will then insert them into the bullet collision pipeline.
- Hopefully forward dynamics will take off correctly after that.
- Sept 15th
- Add callback for broadphase and narrowphase collision detection.
- May be possible to add contact points detected during raytracing overlaps.
- Wrote some tcl code to generate scenes automatically and raytrace them.
- Sept 7th - Sept 14th
- Chugging along, now added code to retrieve bounding boxes from Bullet
- Now object activation states inside Bullet (idle or sleeping) are displayed through different colors on the boxes
- Wrote some initial code to shoot rays and add points to the bullet points cache.
- Sept 4th - Sept 6th
- Not much coding done, was mostly going through the collision detection code in Bullet
- Almost done adding bounding box and object state colors for easier debugging in mged.
- Apparently a solution to arbitrary shapes is to remove all contact manifolds generated by Bullet and apply a force ourselves. Need to test the idea.
- Sep 3rd
- Corrected object positioning code in mged
- Sep 2nd
- Updated the code for the simulate command. Now objects fall "out of the sky"
- Finally reached the first milestone : http://www.youtube.com/watch?v=SByoQQStH2s
- Aug 29th - Sep 1st
- Added regions and groups to the bounding box function
- Added code to pass regions from a model to bullet for the sim
- Aug 22nd - Aug 28th
- Finished the bounding box function
- Modified CMake logic to compile the simulate command as a separate library
- Working on passing multiple objects from a Mged model to the simulation so scenarios like n balls on a billiard table can be simulated.
- Aug 21st
- Compiled Bullet with libged and loaded its dynamic libs with mged.
- Ran a sphere dropping on a plane scenario and it gave the correct results
- Now updating it to translate the position of a sph in mged to the final position of the sphere, before passing the bounding box etc.
- Aug 13th - Aug 20th
- Submitted a fix for Mged. The issue was that a body may be in edit mode when it killed. Then accepting any tranformations done on it can result in a Segmentation fault due to the NULL illump pointer. titles.c was changed.
- Committed changes for the simulate command in Mged.
- Modified top level CMakelists.txt to link to Bullet libs.
- Had an issue accessing the Bullet headers, but that was due to a non-standard install location of Bullet. Now trying to link to Bullet libs.
- Tried using db_functree() to load missing primitives into the in-mem database instance, but the tree representing the primitives are not available, so trying to find a way to convert the rt_db_internal to union tree.
- More code reading.
- Aug 12th
- Mostly read the extra articles in the BRL-CAD documentation
- Tried to extend the bounding box code to use rt_comb_internal instead of rt_gettree()
- Aug 11th:
- Submitted patch 3390331 for BB of primitives.
- Aug 10th:
- Finished the code for the bounding box. It works now for basics shapes.Should be extended to work for regions and groups.
- Aug 9th:
- Mostly reading libwdb and librt code and finishing the function
- Read some of the tutorial articles on BRL-CAD
- Aug 8th:
- Mostly reading code and documentation.
- Started working on function for librt which calculates the bounding RPP when a shapes, region or group is passed in the database internal format struct rt_db_internal.
- Aug 7th:
- Figured out how to transform objects to their new orientations and positions now in Mged !. The plan is to now move geometry in a model using the output transformation matrices of a physics sim. However mged draws the object only when the command is finished, so the command needs to be repeatedly called for multiple steps, to see the object animate.
- Changed the runphysics command to "simulate" to imply more general usage based on brlcad's suggestion.
- Working on a wrapper for the C++ based Bullet physics engine code to be called from the simulate.c file.
- Started maintaining daily logs
- Aug 5th - Aug 6th:
- Inserted a runphysics command in Mged.
- Working on moving objects using matrix transforms inside Mged
- Reading principles of effective modeling and going through code in librt.
- Downloaded and compiled Bullet to make sure its installed and working.
- Aug 3rd - Aug 4th:
- Finished most of the tutorials and articles for BRL-CAD
- Worked through the Interactive Ray Tracing using NIRT guide to get familiar with ray tracing.
- Converted the bounding box program to use librt functions only.
- Aug 1st - Aug 2nd:
- Compiled BRL-CAD on OpenSUSE 11.4 using debugging options
- Went through code in libged especially get_obj_bounds.c
- Finished a stand alone program based on the very high level libged for getting the bounding box of arbitrarily shaped geometry.
Updated Development Time line(Oct 6th)
- Week 6 : Integrate rt and Bullet for detecting overlaps
- Retrieve/display bounding boxes, object states.
- Get the ray tracer integrated so basic spheres can collide and roll across a surface, basically rays have to be shot so that points inside overlapping regions can be added to the bullet collision points cache.
- Add flags to control bounding box display and object states. Also planning to allow objects to be shot towards a scene, with a custom force and initial location.
- Week 7 : Work with more objects and larger systems. The basic focus is to realize a billiard table with multi-colored balls rolling across based on the force applied with a stick.
- Week 8 : Continue with fine-tuning the code and improving collision detection accuracy
- Week 9 : Finish the raytrace based manifold generation code.
- Week 10: Continue to improve the manifold generation. Testing will be required for static objects resting on top of each other and thus not overlapping. How will rays detect the overlap in this case ?
- Week 11: Start working on documentation for the simulate command. Integrate the code for applying forces and setting initial mass which had been inserted before but removed for debugging reasons temporarily. Significant chunks of code need to be converted from libged to librt for performance reasons.
- Week 12: Finish documentation and cleanup code. Present results of past 3 months and plan the improvements to the code.
Expected Result and eventual benefits for BRL-CAD
The expected result is to allow the user to model an object such as a sphere or a box and apply forces on them. The forces can be gravity, but will also allow the user to specify other additional forces to be applied during the simulation. The user can then “turn on” the physics using a simple command. The command allows the user to specify the number of physics steps as well. After the simulation begins the user will be able to see the object move in the GUI as the physics is updated. As time permits accurate material properties can be allowed to be specified and will be used to affect object collisions during the simulation. The system will allow frames to be captured during the simulation for video playback later on. Adding accurate physics to BRL-CAD will expand the scope of its usage allowing users to simulate and test concepts within the software. It provides an exciting addition to a powerful piece of software.
- Bullet Physics: http://bulletphysics.org/mediawiki-1.5.8/index.php/Bullet_User_Manual_and_API_documentation
- Bullet custom collision callbacks and triggers : http://bulletphysics.org/mediawiki-1.5.8/index.php/Collision_Callbacks_and_Triggers
- Bullet User Manual: http://code.google.com/p/bullet/source/browse/trunk/Bullet_User_Manual.pdf
- Open Dynamics Engine: http://www.ode.org/