RAFL: Generalizable Sim-to-Real of Soft Robots with Residual Acceleration Field Learning

Dong Heon Cho*, Boyuan Chen

Duke University

Project Website | Video | Paper

Overview

This repo contains the PyTorch implementation for our paper "RAFL: Generalizable Sim-to-Real of Soft Robots with Residual Acceleration Field Learning".

Note: this is a modified version of https://github.com/srl-ethz/residual_physics_sim2real.git and https://github.com/mit-gfx/diff_pd, with the purpose of learning generalizable residual acceleration fields of soft deformable objects.

Citation

If you find our paper or codebase helpful, please consider citing:

@misc{cho2026raflgeneralizablesimtorealsoft,
      title={RAFL: Generalizable Sim-to-Real of Soft Robots with Residual Acceleration Field Learning}, 
      author={Dong Heon Cho and Boyuan Chen},
      year={2026},
      eprint={2603.22039},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2603.22039}, 
}

Recommended systems

• Ubuntu 18.04
• (Mini)conda 4.7.12 or higher
• GCC 7.5 (Other versions might work but we tested the codebase with 7.5 only)

Content

• Installation
• Sim-to-Sim
• Sim-to-Real
• License

Installation

Recommended systems

• Ubuntu 18.04
• (Mini)conda 4.7.12 or higher
• GCC 7.5 (Other versions might work but we tested the codebase with 7.5 only)

Clone the repository and install the dependencies.

git clone --recursive https://github.com/generalroboticslab/RAFL.git
cd RAFL
conda env create -f environment.yml
conda activate residual_physics
./install.sh

Sim-to-Sim

Navigate to the corresponding sub-directory by

cd python/beam_model/sim2sim_beam_model/[vibration/twist]

Data Generation

To generate the target simulation trajectories and optimized residual forces for the canonical beam

python data_generation.py
python optimize_trajectories.py

This will additionally generate target residual forces for the supervised residual learning baseline. All raw trajectories will be stored under data_real subdirectory and all target datasets will be stored under data_sim2sim.

Addtionally, to generate target simulation trajectories for various shapes run

python data_generation_[shape].py
python optimize_trajectories_[shape].py

where shape can be the following

• longer: Longer beam (unscaled)
• longer_scaled: Longer beam (scaled)
• shorter: Shorter beam (unscaled)
• shorter_scaled: Shorter beam (scaled)
• thicker: Thicker beam (unscaled)
• thicker_scaled: Thicker beam (scaled)
• thinner: Thinner beam (unscaled)
• thinner_scaled: Thinner beam (scaled)
• fishTail: Crude Fishtail
• fishTailMesh: Finer Fishtail

Target residual forces for the supervised residual learning baseline are NOT generated for generalization datasets, but target datasets are reformatted to allow compatibility with test code. All raw trajectories will be stored under [shape]_data_real subdirectory and all target datasets will be stored under [shape]_data_sim2sim.

Training

To train our RAFL framework directly on target trajectories, run

python training_direct.py

The trained model will be stored under training/test_refactor_element_direct subdirectory

To train the supervised residual baseline, run

python training.py

The trained model will be stored under training/test_refactor subdirectory

Testing

To test the model, run

python test_residual_physics.py

Resulting trajectory plots are stored under sim2sim/test_refactor_element_direct subdirectory and raw trajectory/error data are stored under the training/test_refactor_element_direct subdirectory

For generalization test, run

python test_residual_physics_[shape].py

Resulting trajectory plots are stored under [shape]_sim2sim/test_refactor_element_direct subdirectory and raw trajectory/error data are stored under the training/test_refactor_element_direct/[shape] subdirectory

Fine-tuning

For the two fish tail shapes, we additionally allow further finetuning via the function

python fine_tune_[fishTail/fishTailMesh].py

The finetuned model will be stored under ``training/test_refactor_element_direct/[fishTail/fishTailMesh]_finetune``` subdirectory

To test the finetuned models run

python test_residual_physics_[fishTail/fishTailMesh]_finetune.py

Resulting trajectory plots are stored under [fishTail/fishTailMesh]_finetune_sim2sim/test_refactor_element_direct/[fishTail/fishTailMesh]_finetune subdirectory and raw trajectory/error data are stored under the training/test_refactor_element_direct/[fishTail/fishTailMesh]_finetune subdirectory

Sim-to-Real

Navigate to the corresponding sub-directory by

cd python/beam_model/sim2real_beam_model

Data Generation

You can download Sim2real beam real data from Google Drive , where data_new, data_longer, data_shorter, data_thicker, data_thinner store the collected raw data and cantilver_data_new_straight, cantilver_data_longer_straight, cantilver_data_longer_scaled_straight, cantilver_data_shorter_straight, cantilver_data_shorter_scaled_straight, cantilver_data_thicker_straight, cantilver_data_thicker_scaled_straight, cantilver_data_thinner_straight, cantilver_data_thinner_scaled_straight store the target datasets with optimized initial states and target forces where applicable.

Alternatively, to generate the initial states and optimized residual forces for the canonical beam from raw data, run

python generate_augmented_data.py

This will additionally generate target residual forces for the supervised residual learning baseline. All target datasets will be stored under cantilever_data_new_straight.

Addtionally, to generate target simulation trajectories for various shapes run

python generate_data_[shape].py

where shape can be the following

• longer: Longer beam (unscaled)
• longer_scaled: Longer beam (scaled)
• shorter: Shorter beam (unscaled)
• shorter_scaled: Shorter beam (scaled)
• thicker: Thicker beam (unscaled)
• thicker_scaled: Thicker beam (scaled)
• thinner: Thinner beam (unscaled)
• thinner_scaled: Thinner beam (scaled)

Target residual forces for the supervised residual learning baseline are NOT generated for generalization datasets, and only initial states are generated. All target datasets will be stored under cantilver_data_[shape]_straight.

Training

To train our RAFL framework directly on target trajectories, run

python training_direct.py

The trained model will be stored under training/test_refactor_element_direct subdirectory

To train the supervised residual baseline, run

python training.py

The trained model will be stored under training/test_refactor subdirectory

Testing

To test the model, run

python test_residual_physics.py

Resulting trajectory plots are stored under 2dplots_displacement/test_refactor_element_direct subdirectory and raw trajectory/error data are stored under the training/test_refactor_element_direct subdirectory

For generalization test, run

python test_residual_physics_[shape].py

Resulting trajectory plots are stored under [shape]_2dplots_displacement/test_refactor_element_direct subdirectory and raw trajectory/error data are stored under the training/test_refactor_element_direct/[shape] subdirectory

Expanded Generalization

Additionally, RAFL can be trained on different shapes without requiring target residual forces.

To train on each shape, run

python training_direct_[shape].py

The trained model will be stored under training/test_refactor_element_[shape]_direct subdirectory

System Identification

To run system identification (fitting both Young's Modulus and Poissons's Ratio) on the canonical beam, run

python beam_sys_all.py

and for each shape,

python beam_sys_all_[shape].py

License

This repository is released under the MIT license. See LICENSE for additional details.