DLO-Lab: Benchmarking Deformable Linear Object Manipulations with Differentiable Physics

Junyi Cao · Yian Wang · Ziyan Xiong · Chunru Lin · Zhehuan Chen · Chuang Gan

Paper    |    Website

DLO-Lab supports various material properties for deformable linear objects and their interactions with rigid and soft bodies, allowing for versatile manipulations.

Please consider citing our paper if you find it interesting or helpful to your research.

@inproceedings{Cao_2026_DLOLab,
    author      = {Cao, Junyi and Wang, Yian and Xiong, Ziyan and Lin, Chunru and Chen, Zhehuan and Gan, Chuang},
    title       = {DLO-Lab: Benchmarking Deformable Linear Object Manipulations with Differentiable Physics},
    booktitle   = {International Conference on Machine Learning (ICML)},
    year        = {2026}
}

---

Installation

1. Create conda environment

conda create -n dlolab python=3.12
conda activate dlolab

2. Install PyTorch following the official instructions

3. Install DLO-Lab via PyPI

git clone https://github.com/UMass-Embodied-AGI/DLO-Lab.git
cd DLO-Lab
pip install -e ".[dlo-lab,dev]"

4. Install ffmpeg

conda install "ffmpeg=7" -c conda-forge

5. Install Mushroom-RL

git clone https://github.com/XJay18/mushroom-rl.git
cd mushroom-rl
pip install -e .

6. Download DLO-Lab assets from this link, unzip it under genesis/assets, make sure you have this folder: genesis/assets/dlo-lab

---

Quick Example

DLO-Lab is built on top of Genesis. A deformable linear object is added to a scene as an entity with a ROD material and a Rod/ParameterizedRod morph, then simulated by stepping the scene. The snippet below simulates two ropes (each with one end fixed) colliding with two fixed capsules:

import genesis as gs

gs.init(seed=0, precision="64", logging_level="info", backend=gs.gpu)

# Create a scene with the rod solver activated
scene = gs.Scene(
    sim_options=gs.options.SimOptions(dt=1e-3, substeps=5),
    rod_options=gs.options.RODOptions(damping=10.0, angular_damping=5.0),
    show_viewer=True,
)

# Ground plane
scene.add_entity(
    material=gs.materials.Rigid(needs_coup=True, coup_friction=0.1),
    morph=gs.morphs.Plane(fixed=True),
)

# Two thin ropes. E/G control bending/twisting stiffness
v1 = scene.add_entity(
    material=gs.materials.ROD.Base(segment_radius=0.005, E=1e5, G=1e4),
    morph=gs.morphs.ParameterizedRod(
        type="rod", n_vertices=100, interval=0.01, axis="x",
        pos=(0.5, 0.5, 0.3), euler=(0.0, 0.0, 15.0),
    ),
)
v2 = scene.add_entity(
    material=gs.materials.ROD.Base(segment_radius=0.005, E=1e5, G=1e4),
    morph=gs.morphs.ParameterizedRod(
        type="rod", n_vertices=80, interval=0.01, axis="x",
        pos=(0.55, 0.43, 0.4), euler=(0.0, 0.0, 0.0),
    ),
)

# Two thick, short capsules for the ropes to collide with
b1 = scene.add_entity(
    material=gs.materials.ROD.Base(segment_radius=0.02),
    morph=gs.morphs.ParameterizedRod(
        type="rod", n_vertices=3, interval=0.1, axis="x",
        pos=(0.75, 0.435, 0.25), euler=(0.0, 0.0, -75.0),
    ),
)
b2 = scene.add_entity(
    material=gs.materials.ROD.Base(segment_radius=0.02),
    morph=gs.morphs.ParameterizedRod(
        type="rod", n_vertices=3, interval=0.1, axis="x",
        pos=(1.05, 0.435, 0.25), euler=(0.0, 0.0, -75.0),
    ),
)

# Build, then pin selected vertices: one end of each rope, and the capsules
scene.build(n_envs=1)
v1.set_fixed_states(fixed_ids=[0, 1])
v2.set_fixed_states(fixed_ids=[78, 79])
b1.set_fixed_states(fixed_ids=[0, 1, 2])
b2.set_fixed_states(fixed_ids=[0, 1, 2])

# Roll out the simulation
for _ in range(2000):
    scene.step()

For complete, runnable scripts — grasping with a Franka arm, coupling rods with elastic/sand/liquid bodies, and differentiable system identification — see examples/.

# Run the quick example, saving a video
python examples/quick_example.py --output_folder ./output

---

DLO-Lab Benchmark

DLO-Lab benchmarks eight manipulation tasks — coiling, gathering, lifting, separation, slingshot, unknotting, wiring_post, wrapping — across several learning and trajectory-optimization methods. You could also extend beyond these tasks with the provided APIs.

Note: All commands below are run from the experiments/ directory.

Training

Ready-to-run launch scripts live in experiments/scripts/, organized one folder per method, one script per task:

Method	Scripts	Entry point
PPO	scripts/ppo/	rl/rudinppo.py
SAC	scripts/sac/	rl/sac.py
SHAC	scripts/shac/	rl/shac.py
SAPO	scripts/sapo/	rl/shac.py
CMA-ES	scripts/cmaes/	trajopt/cmaes.py
Gradient Descent	scripts/gd/	trajopt/gd.py

For example, to train PPO on the coiling task:

cd experiments
bash scripts/ppo/coiling.sh

Logs and checkpoints are written to logs/<task>/<exp_name>/, where <exp_name> is set by the --exp_name flag inside each script (e.g. rudin-01 for PPO, sac-01 for SAC, cmaes-01 for CMA-ES).

Visualizing a trained policy (test)

To roll out and render a trained policy/trajectory, re-run the entry point directly with the same --task and --exp_name, plus the method-specific test flag. This evaluates the checkpoint and saves an animation to logs/<task>/<exp_name>/. Add -r for ray-traced rendering, or --gui for the interactive viewer.

# RL methods — load a checkpoint with --test
python rl/rudinppo.py --task coiling --exp_name rudin-01 --test best   # PPO: "best" or an epoch number
python rl/sac.py      --task coiling --exp_name sac-01   --test best   # SAC: "best" or an epoch number
python rl/shac.py     --task coiling --exp_name shac-01  --test best_shac.pkl   # SHAC/SAPO: checkpoint filename

# Trajectory optimization — replay the optimized trajectory
python trajopt/cmaes.py --task coiling --n_steps $n_steps --exp_name cmaes-01 --vis_traj logs/coiling/cmaes-01/best_qpos.npy  # $n_steps should match the training script
python trajopt/gd.py    --task coiling --n_steps $n_steps --exp_name gd-01    --test best_traj.pt # $n_steps should match the training script

RL checkpoints are saved under logs/<task>/<exp_name>/ckpts/; --test best loads the best-return checkpoint, while --test <epoch> loads a specific one. For trajectory optimization, point --vis_traj / --test at the saved trajectory file under the run's log directory.

Ray-traced rendering

If you want to use the ray-traced renderer, please follow the instruction to install LuisaRenderer first.

Adding a new task

Please follow this guide for adding a new manipulation environment.

Simulation demo collection

We offer APIs for collecting simulation demos and saving them as LeRobot v3.0 datasets, which can be used to train visuomotor policies. Please follow this guide for details.

---

Please feel free to contact Junyi Cao (xjay2018@gmail.com) if you have any questions about this work.