Accepted to IROS 2026

HJCD-IK: GPU-Accelerated Inverse Kinematics
through Batched Hybrid Jacobian Coordinate Descent

1Columbia University  ·  2Purdue University  ·  3Barnard College, Columbia University & Dartmouth College

Real-time, collision-free IK tracking a target amid moving obstacles on a 7-DoF Franka Research 3.

Abstract

Inverse Kinematics (IK) is a core problem in robotics, in which joint configurations are found to achieve a (collision free) desired end-effector pose. Modern IK solvers face a fundamental trade-off: analytical methods are fast but lack generality, while numerical optimization-based methods are broadly applicable but prone to local minima and high computational costs. To overcome this challenge, we introduce HJCD-IK, a GPU-accelerated, sampling-based hybrid solver. By pairing a novel orientation-aware greedy coordinate descent initialization with Jacobian-based polishing and a parallel collision filter, our method achieves up to order-of-magnitude gains in speed and accuracy over state-of-the-art solvers, consistently finding collision-free solutions on the accuracy-latency Pareto frontier, while producing a diverse distribution of high-quality samples. We validate our solver on a physical Franka manipulator and release our code open-source.

Approach

HJCD-IK is a GPU-accelerated, three-phase sampling-based solver — co-designed for block-, warp-, and thread-level parallelism.

HJCD-IK pipeline: orientation-aware CCD initialization, parallel Jacobian polishing, and a parallel collision filter.

1 · Orientation-aware CCD (PO-CCD)

Hundreds to low-thousands of seeds run in parallel. A novel orientation-aware greedy coordinate descent — projecting in both position and orientation space — produces fast, diverse coarse solutions from random initializations.

2 · Parallel Jacobian IK (PJ-IK)

The best seeds are polished by a batched Levenberg–Marquardt solver (trust region + line search, with dogleg / single-coordinate fallbacks), built on the GRiD library, to sub-millimeter and sub-degree accuracy.

3 · Parallel collision filter

A GPU-parallel two-stage hierarchical collision check against bounding spheres filters the refined batch for feasibility, returning collision-free solutions on the self-motion manifold.

Results

Distribution of collision-free IK solutions for cuRobo, PyRoki, and HJCD-IK at a representative target.

Distribution of collision-free IK solutions for a representative target — cuRobo (left), PyRoki (center), HJCD-IK (right). HJCD-IK returns a broader spread of locally-optimal solutions, and a more accurate final result, in less time.

We benchmark HJCD-IK against three GPU-parallel solvers — cuRobo, IKFlow, and PyRoki — on the 7-DoF Franka Panda and Fetch arms (100 Halton open-world poses; the box_panda MotionBenchMaker scene for collision-free), on an NVIDIA RTX 4060. HJCD-IK stays on or near the accuracy–latency Pareto frontier across every batch size and DoF, with order-of-magnitude gains:

Solve time vs. DoF (ms, B = 1000)

DoFHJCD-IKPyRokicuRobo
7 4.2515.099.11
124.5516.2912.66
184.6220.8216.26
244.6624.3419.55

Lowest error at every DoF, too (see paper Table III).

Solution diversity — MMD vs. TRAC-IK (↓)

MetricHJCD-IKPyRokicuRoboIKFlow
MMD 0.022610.045140.053480.03670
MMD²0.000510.002030.002860.00134

Closest match to the full IK manifold.

Accuracy–latency Pareto frontier

Position / orientation / combined error vs. solve time — HJCD-IK (orange), cuRobo (blue), PyRoki (green). HJCD-IK stays on or near the frontier across batch sizes, collision-free problems, and degrees of freedom.

Open-world batch-size accuracy-latency Pareto plot

Open-world, by batch size (Table I)

Collision-free accuracy-latency Pareto plot

Collision-free, box_panda (Fig. 4, Table II)

DoF scaling accuracy-latency Pareto plot

DoF scaling, 7–24 (Fig. 5, Table III)

Full per-batch open-world and collision-free tables (Tables I–II) are in the documentation.

Hardware deployment

HJCD-IK running on a 7-DoF Franka Research 3, tracking a target around obstacles in real time.

We deploy HJCD-IK on a 7-DoF Franka Research 3. Each camera frame yields the target and obstacle poses; HJCD-IK solves with a batch size of 1000 to obtain a diverse set of collision-free configurations on the self-motion manifold, and we move toward the one nearest the current configuration for smooth, real-time, collision-free tracking of moving obstacles.

BibTeX

@inproceedings{yasutake2026hjcdik,
  title     = {{HJCD-IK}: {GPU}-Accelerated Inverse Kinematics through Batched Hybrid Jacobian Coordinate Descent},
  author    = {Yasutake, Cael and Liu, Andrew H. and Kingston, Zachary and Plancher, Brian},
  booktitle = {2026 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  year      = {2026},
  note      = {arXiv:2510.07514}
}