HJCD-IK#
Hybrid Jacobian Coordinate Descent Inverse Kinematics — a GPU-accelerated, batched inverse kinematics solver that generates many candidate solutions in parallel for a 6-DOF end-effector target, with optional collision avoidance. Each call solves one target: coarse search uses one block per candidate, then Levenberg–Marquardt refinement uses one warp per candidate.
Built on GRiD (robot kinematics codegen) and GLASS (single-block / warp-scoped CUDA linear algebra).
Paper: arXiv:2510.07514 (IROS 2026). Project page: a2r-lab.org/HJCD-IK.
Install the extension and run your first batched IK solve.
The three-phase algorithm: coordinate-descent init, Jacobian polish, collision filter.
Runnable examples, benchmark results vs. cuRobo / PyRoki / IKFlow, and how to reproduce.
Generate kinematics for a new URDF / end-effector frame.
The Python API and the Doxygen-generated C++/CUDA reference.
Contributing, the codegen/build discipline, and how to edit these docs.
Quick start#
import hjcdik
solver = hjcdik.Solver()
target = solver.sample_targets(1, seed=0)[0] # [x, y, z, qw, qx, qy, qz]
out = solver.solve(target, batch_size=2000, num_solutions=4, collision=None)
r = out[0]
print(r.count, "candidates")
if r.count:
print("candidate position errors (mm):", r.pos_err_mm)
print("candidate orientation errors (rad):", r.ori_err_rad)
print("accurate candidates:", r.q[r.ok(pos_mm=5, ori_rad=0.001)])