Optimization-Based Motion Planning

OPALM: A Cross-Layer Benchmark for Motion Planning

A controlled benchmark for measuring how planner, collision, kinematics, and step-solver choices shape robotic manipulation.

Overview View figures

Overview

Benchmark the whole planning stack.

OPALM separates trajectory formulation, collision geometry, kinematics, and nonlinear solving under a common planning and measurement contract. Five planners share matched initialization, collision queries, work budgets, timing, and dense validation.

  • Five planner adapters: CHOMP, TrajOpt, GPMP2, FACTO, and DRAFTO.
  • Three collision representations with Pinocchio and SIMD kinematics.
  • Comparable success, latency, iterations, roughness, and path length.
OPALM framework showing collision, kinematics, planner, and step solver modules.
Interchangeable components with a shared experimental contract.

Benchmark Design

A matched stack, not a planner-only comparison.

Every admissible run holds the robot, scene, start and goal states, safety margins, evaluation budget, and final verifier fixed.

Planner

Five formulations

Waypoint, factor-graph, and coefficient-space optimizers are compared through native adapters.

Collision

Three representations

SDF, primitive shapes, and CAPT return shared distance and gradient semantics.

Kinematics

Scalar and SIMD

Pinocchio is compared with an AVX, structure-of-arrays execution path.

Step Solver

Matched local models

GN, LM, Dogleg, GD, and SQP isolate solver behavior from objective construction.

Task regimes: S/U single-arm unconstrained, S/C single-arm constrained, D/U dual-arm unconstrained, and D/C dual-arm constrained.

Key Results

Three practical takeaways.

>94%

Coefficient planners

FACTO and DRAFTO exceed 94% success in task-unconstrained settings and remain above 90% under the studied constraints.

14.5x

Forward kinematics

Median SIMD acceleration for forward kinematics; the end-to-end benefit depends on the planner's remaining bottleneck.

GN / LM

Solver choice matters

GN is DRAFTO's fastest option; LM is the more robust GPMP default and improves some dual-arm cases.

Results Figures

Evidence by study.

Each tab keeps the relevant plots together without turning the page into a long plot stream.

Cross-scene distributions for planning time, iterations, roughness, and path length across four task regimes and two kinematics backends.
Cross-scene metricsPlanning time, iterations, roughness, and path length for the five default planner stacks.
Success-rate heatmap for the five planners across four task regimes.
Validated successSuccess rate under each planner's native step-solver setting.

Planning Scenes

Representative single- and dual-arm tasks.

Franka Panda arm in the bookshelf thin single-arm planning scene.
Bookshelf thin · S/U
Franka Panda arm in the kitchen constrained planning scene.
Kitchen · S/C
Franka Panda arm in the table-under-pick planning scene.
Table under pick · S/U, S/C
Dual Franka Panda arms in the table-under-pick planning scene.
Table under pick dual · D/U, D/C