Benchmarking Data‐Driven Control of Octopus‐Inspired Soft Arms in Underwater Environment

Octopus‐inspired tendon‐driven soft arms remain difficult to control in underwater environment, despite over a decade of research. Dominant data‐driven approaches—reinforcement and imitation learning—demand hundreds of thousands to millions of training samples, rely on task‐specific pipelines, and are rarely evaluated against shared baselines, making reproducibility and cross‐study comparison practically difficult. We demonstrate that careful design of data collection strategy and input‐feature space alone yields a sample‐efficient, generalizable, and reproducible supervised policy. We present DS‐Adjust , a parameter‐randomized data collection algorithm that produces smooth, jerk‐bounded actuation spanning a continuum of velocity regimes on physical robot in minutes, eliminating the need for simulation frameworks. A policy trained once generalizes zero‐shot across target and path tracking tasks diversity, active hydrodynamic disturbance, frequency‐based domain‐shift, and stochasticity, achieving errors of ≈1.3 %– of the characteristic arm length across all tested scenarios. The pipeline is validated across three distinct tendon‐driven arms, both underwater and in‐air. An ablation study establishes that explicit action history, rather than positional state observations, is the decisive input design choice for capturing memory‐dependent viscoelastic dynamics of these robots. Benchmarking against in‐house and external similar solutions confirms competitive accuracy at a sample cost two to three orders of magnitude lower than comparable data‐driven solutions.