Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Learning to Control Free-Form Soft Swimmers
Authors: Changyu Hu, Yanke Qu, Qiuan Yang, Xiaoyu Xiong, Kui Wu, Wei Li, Tao Du
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our method on a wide range of morphologies, from bio-inspired to unconventional. From this general framework, high-performance swimming patterns emerge that qualitatively reproduce canonical gaits observed in nature without requiring domain-specific priors, where state-of-the-art baselines often fail, particularly on complex topologies like a torus. We evaluate our framework on a diverse set of 3D soft swimmer morphologies, from bio-inspired fish to unconventional morphologies (Fig. 2), demonstrating universal applicability. Our method achieves observable movement patterns in the majority of tested models in forward swimming task, achieving a 50% higher success rate in learning effective swimming gaits compared to state-of-the-art baselines (Wang et al., 2023a), which struggle to produce meaningful motion for the majority of tested morphologies. |
| Researcher Affiliation | Collaboration | 1 Tsinghua University 2 Shanghai Qi Zhi Institute 3 LIGHTSPEED 4 Shanghai Jiao Tong University |
| Pseudocode | No | The paper describes methods through prose and mathematical formulations rather than structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | The code is available at https://github.com/changyu-hu/Free Flow. We open-source our full codebase including the GPU-accelerated simulator and training pipelines at https://github.com/changyu-hu/Free Flow. |
| Open Datasets | No | We construct a novel collection of 12 soft swimmer morphologies (Fig. 2), including 6 bio-inspired and 6 abstract morphologies. |
| Dataset Splits | No | The paper discusses the morphologies used for experiments but does not explicitly provide details about training, validation, or test dataset splits in terms of percentages or sample counts. |
| Hardware Specification | No | The paper mentions a GPU-accelerated simulator for training policies but does not specify the exact GPU model or other hardware components used for the experiments. |
| Software Dependencies | No | The paper mentions several methods and tools like the Lattice Boltzmann Method (HOME-LBM by Li et al., 2023) and f Tet Wild (Hu et al., 2020) but does not provide specific version numbers for these or other software dependencies. |
| Experiment Setup | Yes | We train all tasks with the soft actor-critic (SAC) method (Haarnoja et al., 2018), a widely adopted deep reinforcement learning (DRL) method known for its stability, sample efficiency, and ability to handle continuous action spaces effectively. The reward is defined as R = Rtask + λsmoothpsmooth + λregpreg, Rtask = vmean d, psmooth = ||a alast||2 2/(6m), preg = ||a||2 2/(6m). It consists of three components: a task-specific term Rtask, a smoothness term psmooth with coefficient λsmooth that encourages natural actions, and a regularization term preg with coefficient λreg that penalizes redundant actions. We trained policies for a fish-like swimmer in 2D across six independent runs with different random seeds while keeping all hyperparameters constant. |