Why ik_bench — one Python script benchmarks C++-only solvers

Inverse-kinematics solvers live in an awkward split. The best ones are C++ MoveIt plugins, and some of the most interesting — bio_ik (evolutionary/memetic) and pick_ik (gradient-descent + memetic global) — are not packaged at all: no pip, no conda, no apt. Others (KDL, TRAC-IK) are packaged C++. And there is always someone's pure-Python Jacobian solver. Benchmarking these against each other normally means standing up a C++ test harness, launch files and a parameter server per solver — enough friction that the comparison rarely gets done honestly.

ik_bench does it in one Python file. cppyy + moveit_kit load each C++ plugin in-process through MoveIt's own pluginlib mechanism and call RobotState::setFromIK; the two unpackaged solvers are built from source once and discovered by the same lookup-by-name path as the packaged ones — cppyy never parses a line of their headers, pluginlib just dlopens the compiled .so. The pure-Python baseline is NumPy only. python ik_bench/run_bench.py runs all five on the same Panda, the same seeded targets and the same tolerances, and prints one table (solve-rate, verified success %, accuracy, near-limit behaviour). The punchline: the fastest solver in the table, bio_ik, is one you cannot install — it exists only as C++ source, yet here it is benchmarked, configured and beaten-or-beating its packaged peers from a single script. That is the new use case cppyy_kit unlocks: cppyy as the harness that makes C++-only libraries first-class citizens of a Python benchmark.