Batched IK

Batched IK#

The main reason to use mink-warp over Mink is parallel IK for many worlds with one model and one solver launch grid.

Configuration#

cfg = mw.Configuration(model, nworld=512, device="cuda")
q0 = np.tile(model.qpos0, (512, 1)).astype(np.float32)
# Optional: per-world perturbation
for i in range(512):
    q0[i, 0] += 0.03 * np.sin(i * 2 * np.pi / 512)
cfg.update(q=q0)

All subsequent FK, Jacobians, and solves use shape (nworld, …).

Per-world targets#

Upload a batch of targets once (or only when they change):

import warp as wp

targets = wp.array(poses_np, dtype=float)  # (nworld, 7) wxyz_xyz
frame.set_target(targets)

For a single shared target, pass one SE3 or a length-1 array; tasks broadcast as needed.

Solver loop#

solver = mw.DLSSolver(cfg)
while running:
    solver.solve_and_integrate(tasks, dt=0.01, use_graph=True)
    q_host = cfg.q.numpy()  # boundary copy for visualization

Demos#

Numbered examples under examples/ (increasing complexity):

#

Script

Topic

01

01_panda_ik.py

Basic FrameTask + IKSolver

02

02_constrained_ur5e.py

Hard limits + collision vs environment

03

03_equality_cassie.py

EqualityConstraintTask

04

04_self_collision_dual_iiwa.py

Inter-arm collision avoidance

05

05_relative_frame_g1.py

RelativeFrameTask + humanoid collision

 
uv sync --extra examples
uv run examples/01_panda_ik.py
uv run examples/05_relative_frame_g1.py

Full table: Examples.

Assets live under examples/franka_emika_panda/ and examples/unitree_g1/.

Performance notes#

  • Prefer keeping q, targets, and task buffers on device across steps.

  • Re-upload targets only when they change (see 05_relative_frame_g1.py).

  • Cholesky tile solve cost is ~flat in nworld; FK + Jacobian assembly often dominate at moderate nv (see Benchmarks).