Research Analyzer
← Back ICRA 2026

Differentiable Particle Optimization for Fast Sequential Manipulation

Lucas Chen, Shrutheesh Raman Iyer, Zachary Kingston

PDF

AI summary

Key figure (auto-extracted from paper)
SPaSM achieves millisecond-scale trajectory optimization for sequential manipulation with a 4000× speedup over existing GPU methods while maintaining 100% success rates.
Sequential manipulation GPU acceleration Trajectory optimization Particle optimization Differentiable planning Real-time robotics

Problem

Sequential robot manipulation requires solving high-dimensional, coupled placement and trajectory optimization in real-time, but existing GPU-accelerated methods are bottlenecked by CPU-GPU data transfer overhead and hierarchical decomposition that limits hardware utilization.

Approach

SPaSM uses a fully GPU-parallelized, two-stage particle optimization strategy that jointly optimizes object placements and robot trajectories through end-to-end compiled CUDA kernels, eliminating CPU coordination overhead.

Key results

  • End-to-end GPU-parallelized trajectory optimization without CPU coordination
  • 4000× speedup over cuTAMP with 100% success rate on challenging benchmarks
  • Joint optimization of object placements and robot trajectories to handle kinematic constraints
  • Millisecond-scale solution times enabling real-time reactive manipulation

Why it matters

Enables real-time reactive manipulation and rapid adaptation for dynamic robotic environments, making high-fidelity sequential planning feasible for practical deployment.

Abstract

Sequential robot manipulation tasks require find- ing collision-free trajectories that satisfy geometric con- straints across multiple object interactions in potentially high- dimensional configuration spaces. Solving these problems in real-time and at large scales has remained out of reach due to computational requirements. Recently, GPU-based acceleration has shown promising results, but prior methods achieve limited performance due to CPU-GPU data transfer overhead and complex logic that prevents full hardware utilization. To this end, we present SPaSM (Sampling Particle optimization for Sequential Manipulation), a fully GPU-parallelized framework that compiles constraint evaluation, sampling, and gradient- based optimization into optimized CUDA kernels for end-to-end trajectory optimization without CPU coordination. The method consists of a two-stage particle optimization strategy: first solv- ing placement constraints through massively parallel sampling, then lifting solutions to full trajectory optimization in joint space. Unlike hierarchical approaches, SPaSM jointly optimizes object placements and robot trajectories to handle scenarios where motion feasibility constrains placement options. Exper- imental evaluation on challenging benchmarks demonstrates solution times in the realm of milliseconds with a 100% success rate; a 4000× speedup compared to existing approaches. Code and examples are available at commalab.org/papers/spasm.

Index terms

Motion and Path Planning Task and Motion Planning

Related papers