Hybrik IK Approaches
Summary
Analytic IK approaches yield fast solutions once the right "path to solution" is identified. Not always can those methods be applied to the existing robot, as geometric IK methods require special structures/alignments of joints. To solve the inverse kinematics for arbitrary modular robots (for now, let's assume 6DOF, extension possible) faster, we could leverage methods developed under the keyword "hybrid IK", meaning basically:
- Solving the analytic inverse kinematics for a slightly modified robot that can be dealt with using geometric approaches
- Starting with this initial guess, find a numeric solution to the underlying robot
Knowledge in this domain would also help in "re-using" IK solutions for "similar" modular robots.
Expected behavior
We want a solver that solves the IK for MRRs efficiently by leveraging analytic and numeric approaches.
What's is needed for?
Currently, especially in MRR optimization, the IK is a computational bottleneck. Any speedup here almost proportionally speeds up our learning pipelines.
Relevant logs, links, or screenshots
For the analytic part of this problem, the IKBT library sounds promising, especially since it comes with a certain level of "explainability" that could be leveraged to understand where the numeric optimization should take over. Link: https://github.com/uw-biorobotics/IKBT
Roadmap
- Literature research on hybrid solvers
- Implementation and runtime estimation for analytic solvers considering both, initial setup and inference
- Definition of the analytic/numeric "interface"
- Runtime evaluation and success rate evaluation for the hybrid approach
/cc @JonathanKuelz