Whole-Body Semantic-to-Actuation Grounding of Elephant-Inspired Soft-Trunk Motion via Lightweight Flow Matching

Jul 16, 2026

Close-contact human–robot interaction can use trunk-like continuum manipulators to express diverse, whole-body, shape-based behaviors, but mapping open-vocabulary language to feasible motions is difficult because end-effector representations underspecify expression while direct body-shape specification is too high-dimensional. A whole-body semantic-to-actuation grounding framework is proposed for elephant-inspired soft-trunk motion. Open-vocabulary responses are canonicalized into morphology-aligned intent–intensity tuples, then executable tendon-actuation trajectories are generated via compact Catmull–Rom spline controls sampled by a lightweight rectified-flow generator. Evaluations show large improvements in grounding correctness and faster inference, with higher user satisfaction in a 100-participant physical HRI study.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages8

Reading Time20 min

Abstract

Introduction

Motivation and representation mismatch

Proposed semantic-to-actuation grounding interface

Challenges for executability and distributional generation

Why rectified flow for efficient sampling

Whole-Body Semantic-to-Actuation Grounding of Elephant-Inspired Soft-Trunk Motion via Lightweight Flow Matching Tingcong Liu 1 ,2 , Tongshun Chen3 , Siyi Ma4 , Yuhao Wang3 , Aye Phyu Phyu Aung2 , Ibrahim Alsarraj3 , J. Senthilnath2 , Bo An 1 , Ke Wu3 arXiv :2607 . 11018v1 [ cs .RO] 13 Jul 2026 Abstract—For close-contact human–robot interaction (HRI), trunk-like continuum manipulators offer an attractive physical channel for diverse whole-body shape-based expression. However, directly applying conventional vision–language–action (VLA) grounding to such systems is challenging: tip- or endeffector motion underspecifies whole-body expression, whereas direct whole-body specification is too high-dimensional for semantic planning and difficult to keep feasible. To address this representation mismatch, we propose a whole-body semanticto-actuation grounding framework for elephant-inspired softtrunk HRI via lightweight flow matching. The framework grounds open-vocabulary responses produced by a multimodal large language model (MLLM) into executable tendonactuation trajectories that realize distributed continuum-body motion. It first canonicalizes open-vocabulary responses into bounded morphology-aligned intent–intensity tuples that define executable motion families, and then parameterizes continuous whole-body tendon-actuation profiles with compact Catmull– Rom spline controls. Conditioned on the resulting tuple, a lightweight rectified-flow generator learns a distribution over feasible spline controls, enabling efficient one-to-many sampling of smooth and socially valid trunk-motion realizations. We evaluate the framework through semantic-to-actuation grounding ablations and a 100-participant physical HRI study. In the experiments, the proposed framework improves held-out grounding correctness from 25.0% to 77.2% over a rawresponse dense-regression baseline. Compared with a denoisingdiffusion baseline, rectified flow improves correctness from 71.9% to 77.2% and reduces inference time from 7.86 ms to 4.87 ms, while preserving motion diversity. In the physical HRI study, the proposed framework increases the positive overallsatisfaction rating from 46% to 82% over the audiovisual-only baseline, demonstrating the user-facing benefit of the generated soft-trunk motion channel. Keywords: Soft robotics; Whole-body Semantic-toactuation Grounding; Bio-inspiration; MLLM; HRI. I. INTRODUCTION Close-contact human–robot interaction (HRI) brings the robot body into the user’s interpersonal and physical space, where approach distance, bodily behavior, and touch shape perceived appropriateness, comfort, intent, and safety [1],[2] . Since robot motion can function as a communicative resource beyond speech or screen-based display [3], closecontact HRI motivates a physical extension that can act asan embodied channel for social response. Rigid manipulators offer accurate and repeatable motion, but close-contact social use must also address physical safety, perceived safety, and comfort in proximity [4] . Trunk-like soft continuum manipulators are therefore promising: their continuum bodies can generate non-contact body language through approach, withdrawal, and gesture [5], while their compliant morphol- ogy can deform during light contact [6] . However, the same high-dimensional and deformable whole-body morphology that enables expressive and contact-friendly interaction also changes the grounding problem: the behavior perceived by the user is distributed over the entire continuum body rather than concentrated at a single end effector. This creates a mismatch with the tip-centered action representations commonly used in vision-language-action robot policies [7] . Existing soft-robot VLA-style formulations [8] typically represent robot actions through endeffector displacements or low-level control commands. Such representations are suitable for manipulation tasks that demand precise task-space outcomes, but are insufficient for soft tru

FAQs

What benefits were observed in the physical human–robot interaction study?

Compared with an audiovisual-only baseline, the framework increased the positive overall satisfaction rating from 46% to 82%, indicating user-facing improvement from the generated soft-trunk motion channel.

Jul 16, 2026

Upload by Genevieve