SurGE

Paper | Project Page | Video

Official implementation of "SurGE: Surrogate Gradient-guided Evolution for Co-design of Legged Robots with Parallel Elasticity".

Overview

SurGE jointly optimizes the spring design and control policy of legged robots with parallel elasticity. To recover gradient information despite this non-differentiability, SurGE differentiates a surrogate pipeline, a kinodynamic single-rigid-body model paired with a design-aware policy, and injects the resulting surrogate gradient into CMA-ES through a mean shift with cosine-annealed decay. This converges faster and more reproducibly than pure gradient-based or pure evolutionary search, with the improvement transferring to hardware.

Installation

Dependencies

The code is tested on the following setup:

• Ubuntu 22.04
• Python 3.8
• Isaac Gym (Preview 4)

Environment Setup

1. Clone the repo:

   git clone https://github.com/ARCaD-Lab-UM/surge-codesign.git
   cd surge-codesign

2. Create and activate the conda environment:

   conda env create -f environment.yml
   conda activate codesign

3. Download and install Isaac Gym into this env per NVIDIA's instructions.
4. Install the co-design code:

   pip install -e .

The editable install builds all three packages from src/: surge_codesign, legged_gym, and rsl_rl.

Quick Start

Co-design with SurGE

python scripts/run_surge_codesign.py

Baselines

python scripts/run_gd_codesign.py    # vanilla gradient descent
python scripts/run_cma_codesign.py   # vanilla CMA-ES

Visualization & Analysis

python scripts/collect_landscape.py                          # collect objective landscape
python scripts/plot_landscape.py --policy_id rainbow_v7      # plot landscape
python scripts/collect_gradient_field.py                     # collect gradient field
python scripts/plot_gradient_field.py --grad-magnitude 5     # plot gradient field

Train a Design-Aware Locomotion Policy

Pretrained checkpoints are shipped in checkpoints/. To visualize a pretrained policy:

python scripts/play_policy.py --task hopper --load_pretrained_ckpt

To train a new one:

python scripts/train_policy.py --task hopper --headless

Note

To use a newly trained policy in co-design, set policy_root="logs/<exp_name>" and policy_id="<run_name>" in src/surge_codesign/config.py

Code Structure

The co-design logic lives in src/surge_codesign/, alongside our customized RL framework.

src/surge_codesign/         # core co-design package
  config.py                # configuration
  design_space.py          # design parameters and bounds
  design_objective.py      # design objective
  mups_robot.py            # differentiable Kino-SRB surrogate
  mups_spring.py           # differentiable UPS spring model
  optim_helper.py          # rollout and gradient engine
  data_logger.py           # logging
  vis_helper.py            # plotting
src/legged_gym/            # customized legged-gym (hopper simulation env)
src/rsl_rl/                # customized rsl-rl (RL framework)

Citation

If you find this code useful for your research, please consider citing our paper:

@article{zhuang2026surge,
  title={SurGE: Surrogate Gradient-guided Evolution for Co-design of Legged Robots with Parallel Elasticity},
  author={Yulun Zhuang and Yue Qin and Justin Lu and Zelin Shen and Yichen Wang and Sicheng He and Yanran Ding},
  journal={arXiv preprint arXiv 2606.21866},
  year={2026}
}

Troubleshooting

If you see an error about missing libpython3.8.so.1.0 when importing Isaac Gym, copy it from conda lib to isaacgym bindings, i.e.

cp path_to_conda/envs/codesign/lib/libpython3.8.so.1.0 path_to_isaacgym/python/isaacgym/_bindings/linux-x86_64/