跳转至

文章背景与核心概要

基于柯瑟拉杆理论（Cosserat-rod theory）的降阶模型在软体机器人领域已得到广泛应用，但传统实现方式缺乏对先进刚性机器人中常用的可微分、GPU并行以及面向控制的工作流的支持。为了弥补这一空白，本文作者推出了 SoRoMoX（Soft Robot Models in JAX）——一个完全基于数值计算、支持 JIT 编译的 Python/JAX 框架。SoRoMoX 提供了一个统一且面向控制的接口，支持关节型、分段常应变（PCS）以及变应变模型，能够在 GPU 上直接计算惯性矩阵、重力与弹性力、雅可比矩阵及其导数，从而实现关于状态、输入和参数的端到端可微分特性。

在性能与应用方面，SoRoMoX 实现了突破性的提升：其顺序 CPU 仿真速度比同类先进方法快高达 18.1 倍，而 GPU 并行仿真更是将吞吐量提升了高达 234.6 倍。该框架在静态平衡系统辨识中使标记点均方根误差（RMSE）降低了 66%，在残余力学习中实现了额外的 64% 误差约减，并在计算力矩追踪中将 RMSE 降低了约 500 倍。此外，它还支持控制增益优化（损失降低高达 62%）、基于高阶控制障碍函数的安全受限控制（将峰值接触力严格限制在 5 N 以内），并通过大规模并行仿真，使强化学习策略的训练速度比基于 CPU 的 PyElastica 离散杆基准快 7 倍。

SoRoMoX: Fast, Differentiable, and Parallelizable Soft Robot Models

SoRoMoX: Fast, Differentiable, and Parallelizable Soft Robot Models

arXiv: 2608.06650 [cs.RO]
Submitted: August 6, 2026
Primary Subject: Robotics (cs.RO)
Authors: Maximilian Stölzle, Solange Gribonval, Daniel Feliu-Talegon, Vito Daniele Perfetta, Michele Martini, Chuhan Zhang, Kiwan Wong, Mohammed Tarnini, Anup Teejo Mathew, Federico Renda, Daniela Rus, Cosimo Della Santina

arXiv: 2608.06650 [cs.RO]
Submitted: August 6, 2026
Primary Subject: Robotics (cs.RO)
Authors: Maximilian Stölzle, Solange Gribonval, Daniel Feliu-Talegon, Vito Daniele Perfetta, Michele Martini, Chuhan Zhang, Kiwan Wong, Mohammed Tarnini, Anup Teejo Mathew, Federico Renda, Daniela Rus, Cosimo Della Santina

Abstract Summary

Abstract Summary

While reduced-order models based on Cosserat-rod theory are well-established for soft robotics, traditional implementations lack support for the differentiable, GPU-parallel, and control-oriented workflows commonly used in advanced rigid robotics.

While reduced-order models based on Cosserat-rod theory are well-established for soft robotics, traditional implementations lack support for the differentiable, GPU-parallel, and control-oriented workflows commonly used in advanced rigid robotics.

To bridge this gap, the authors introduce SoRoMoX (Soft Robot Models in JAX)—a fully numerical, JIT-compilable Python/JAX framework. SoRoMoX provides a unified, control-ready interface supporting articulated, Piecewise Constant Strain (PCS), and Variable Strain models. It computes inertia matrices, gravitational and elastic forces, Jacobians, and their derivatives directly on GPUs, making it end-to-end differentiable with respect to states, inputs, and parameters.

To bridge this gap, the authors introduce SoRoMoX (Soft Robot Models in JAX)—a fully numerical, JIT-compilable Python/JAX framework. SoRoMoX provides a unified, control-ready interface supporting articulated, Piecewise Constant Strain (PCS), and Variable Strain models. It computes inertia matrices, gravitational and elastic forces, Jacobians, and their derivatives directly on GPUs, making it end-to-end differentiable with respect to states, inputs, and parameters.

Key Performance & Application Highlights:

  • Massive Speedups: Sequential CPU rollouts are up to 18.1× faster than state-of-the-art alternatives, while GPU-parallel rollouts increase throughput by up to 234.6×.
  • Static-Equilibrium System Identification: Achieves 66% lower marker RMSE.
  • Residual-Force Learning: Delivers an additional 64% reduction in error.
  • Computed-Torque Tracking: Reduces RMSE by a factor of approximately 500 compared to model-free PD controllers.
  • Control-Gain Optimization: Yields up to 62% lower loss than untuned gains.
  • Safety-Constrained Control: Utilizes high-order control barrier functions to constrain peak contact forces within a strict 5 N bound (compared to 33.5 N without constraints).
  • Reinforcement Learning: Trains RL policies up to 7× faster than a CPU PyElastica discrete-rod baseline through massively parallel rollouts.

Key Performance & Application Highlights:

  • Massive Speedups: Sequential CPU rollouts are up to 18.1× faster than state-of-the-art alternatives, while GPU-parallel rollouts increase throughput by up to 234.6×.
  • Static-Equilibrium System Identification: Achieves 66% lower marker RMSE.
  • Residual-Force Learning: Delivers an additional 64% reduction in error.
  • Computed-Torque Tracking: Reduces RMSE by a factor of approximately 500 compared to model-free PD controllers.
  • Control-Gain Optimization: Yields up to 62% lower loss than untuned gains.
  • Safety-Constrained Control: Utilizes high-order control barrier functions to constrain peak contact forces within a strict 5 N bound (compared to 33.5 N without constraints).
  • Reinforcement Learning: Trains RL policies up to 7× faster than a CPU PyElastica discrete-rod baseline through massively parallel rollouts.