Battle-proven inverse kinematics and planners for constrained environments
Hi,
We are currently using PyBullet with custom inverse kinematics and planner. We are wondering if ROS has battle-proven tools fitted to our use case that we shall explore.
Our issue is that we have a robotic arm that has to navigate into a very constrained environment. It is a small box that encapsulates the robot, with lots of obstacles. There is always a free path, but it can be a narrow corridor.
The constraints are:
The trajectories are never straight and can be rather complex.
The inverse kinematics shall converge relatively fast. Currently, we are using iterative "accurate" inverse kinematics, which convergence time can vary significantly even with relatively large thresholds.
The joints-limits shall be respected. PyBullet does not guarantee it. So when the inverse kinematics result is out of limits, we have to recalculate one pushing it back into the limits.
The self-collisions between the robot's links shall be avoided during inverse kinematics. So when it is not, we recalculate inverse kinematics pushing it out of the self-collision, which means that we push the robot to unfold itself.
The collision between intermediary robot links and the environment shall also be avoided. So when it is not, we recalculate the inverse kinematic pushing it of collisions, which means that we push the robot to fold itself. It is pretty much the inverse objective as precedently while both can happen at the same time.
Of course, the collision with the tool and the environment shall be avoided, but this is the easiest part to handle as we can push the tool center point out of the collisions.
This implementation with PyBullet is working in most cases, but in terms of performance, it can be very slow (a few seconds to something like 30s) and there are some bugs that can prevent convergence. We are currently wondering if we shall continue with DIY on PyBullet or if she shall have a look at some more advanced ROS, MooveIt and Gazebo tools that have been tested to work in such use cases by the community.
If you have some experience with such constrained environments, I would appreciate it if you could point me to some ROS tools that worked for you.
Thanks a lot
Just an observation: many of your bullets (or the behaviour of your current implementation) are not often considered the responsibility of "IK solvers" in something like MoveIt. IK solvers solve IK / FK requests. Motion planners check whether they are "valid", whatever that means to those motion planners (ie: collision free or adhering to some other criteria).
There are possibilities for IK solvers to receive a couple of callbacks which would allow them to check for things like collisions, but I'm not sure those are always used.
Regarding your motion planning requirements: apart from the narrow space aspect, all of your bullets are taken care of by default by the planners something like MoveIt is using (OMPL, CHOMP, STOMP, etc). Depending on the exact planner used, narrow passages can be difficult to plan for, but it's doable.
I'd also recommend taking a look at Tesseract. Similar tool, different focus.
I guess you could summarise my comment as: don't over-constrain your search by just looking for IK solvers.
I'd look at motion planners in general. They'll always come with IK / FK solvers. How they interact with those solvers would come down to how the authors see the division of responsibilities.
Thanks. yes, our implementation is currently part in the IK solver and part in a planner. That's more a convention as there are actually 3 layers: the IK planner by itself, the collision avoidance without moving the tool by iterating in the null space, and the collision avoidance by moving the tool.
And yes, we are considering both IK and motion planners.