SoftMimic Code Release

This repository provides an implementation of the paper:

SoftMimic: Learning Compliant Whole-body Control from Examples
Gabriel B. Margolis*, Michelle Wang*, Nolan Fey, and Pulkit Agrawal
preprint, 2025
paper / project page

If you use this repository in your work, consider citing:

@article{margolis2025softmimic,
    title={{SoftMimic}: Learning Compliant Whole-body Control from Examples},
    author={Margolis, Gabriel B. and Wang, Michelle and Fey, Nolan and Agrawal, Pulkit},
    journal={arXiv preprint arXiv:2510.17792},
    year={2025}
}

The project is distributed under the MIT License (see LICENSE). Third-party components such as unitree_sdk2 retain their original licenses in third-party/unitree_sdk2.

Code in unitree_sdk2 for LCM interface comes from OpenHomie: https://github.com/InternRobotics/OpenHomie which is based on walk-these-ways: https://github.com/Improbable-AI/walk-these-ways

This code depends on a number of open-source projects without which the system would not be possible:

• IsaacLab: https://github.com/isaac-sim/IsaacLab
• RSL-RL: https://github.com/leggedrobotics/rsl_rl
• Mink: https://github.com/kevinzakka/mink
• MuJoCo: https://github.com/google-deepmind/mujoco
• unitree_sdk2: https://github.com/unitreerobotics/unitree_sdk2

1. Installation

git clone https://github.com/Improbable-AI/softmimic_release_dev.git
cd softmimic_release_dev
python3 -m venv .venv
source .venv/bin/activate
pip install --upgrade pip
pip install -e ".[all]"  # installs optional extras (mujoco, lcm, mink, seaborn)

Use pip install -e softmimic_gym[isaac] if you intend to run Isaac-based training scripts.

---

2. Repository Layout

softmimic_release_dev/
├── compliant_motion_augmentation/   # Compliant motion augmentation pipeline
├── pretrained_models/               # Exported policies (JIT + params)
├── scripts/                         # RSL-RL training/eval helpers
├── softmimic_deploy/                # Deployment interfaces, sensors, utils
├── softmimic_gym/                   # Isaac Lab extension + packaging
└── third-party/unitree_sdk2/        # Unitree SDK source (for deployment)

---

3. Deployment Interfaces

3.1 MuJoCo simulation

python softmimic_deploy/src/deploy_policy_interface.py \
  --interface mujoco \
  --policy ../../pretrained_models/2025-09-26_03-54-58_StaticStand-SoftMimic/model_48000.jit \
  --motion_path stand.csv \
  --render

Controls (mujoco viewer):

• L: calibrate to the initial pose,
• X: start the policy,
• Y: enter damping mode,
• Arrow keys/PageUp/PageDown: joystick emulation.

The CLI accepts any motion CSV under datasets/motions_csv/ or SOFTMIMIC_DATA_ROOT. Paths outside that tree can be passed explicitly.

3.2 Unitree G1 hardware

1. Network setup

   • Connect via Ethernet and assign a static IP on 192.168.123.0/24.
   • Ensure ssh unitree@ROBOT_IP is reachable (default G1 ROBOT_IP is 192.168.123.164 and default password 123).

2. Repository sync

   # Optional overrides: ROBOT_HOST, TARGET_DIR
   ./send_to_g1.sh

3. On-robot environment

   ssh unitree@192.168.123.164
   cd ~/softmimic_release
   wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-aarch64.sh
   bash Miniconda3-latest-Linux-aarch64.sh
   conda create -n softmimic python=3.10
   conda activate softmimic
   pip install -e ".[lcm]"
   sudo apt-get install liblcm-dev

4. Unitree SDK build

   cd third-party/unitree_sdk2
   mkdir -p build && cd build
   cmake ..
   make -j

5. Deployment

   • Terminal 1 (./g1_control eth0) to bridge between LCM and the Unitree SDK.

   • Terminal 2 to launch the policy:

     python softmimic_deploy/src/deploy_policy_interface.py \
       --interface lcm \
       --policy ../../pretrained_models/2025-09-26_03-54-58_StaticStand-SoftMimic/model_48000.jit

   • Button guide matches the physical remote (L1: initial pose, X: deploy, Y: e-stop/damping mode).

---

4. Training, Evaluation, and Augmentation

4.1 Compliant Motion Augmentation

Visualize the augmentation process in Mujoco:

python mink_generator_ff.py interactive     --motion_path ../datasets/motions_csv/stand.csv     --force_mode forcefield

python mink_generator_ff.py interactive     --motion_path ../datasets/motions_csv/boxpick.csv     --force_mode collision-emulator

Generate the full augmented dataset:

cd compliant_motion_augmentation
./generate_all.sh  # batch-generate datasets listed in config.py

To generate a compliant dataset for a new csv file, simply edit generate_all.sh.

We follow the CSV formatting convention of Unitree's LAFAN1 Retargeting Dataset, appending two columns that annotate the foot contact state.

4.2 Isaac Lab tasks (softmimic_gym)

Install the extension then register environments:

cd softmimic_gym
pip install -e ".[isaac]"
cd ..
python scripts/list_envs.py  # prints registered task IDs

Training example (requires Isaac Sim running or headless):

python scripts/rsl_rl/train.py \
  --task Isaac-G1-AugmentedReference-ForceTorque-Control-VariableStiffness-Hybrid-Deployable-Mimic-v0 --headless

4.3 Policy evaluation and video export

python scripts/rsl_rl/test.py \
  --task Isaac-G1-AugmentedReference-ForceTorque-Control-VariableStiffness-Hybrid-Deployable-Mimic-v0 \
  --num_envs 1 \
  --video --video_length 2000 --headless

---

5. Configuration and Safety Notes

• Follow all safety guidelines given by Unitree.
• Ensure that only one instance of the deployment script runs on the robot at a time.
• This is research code provided as-is to facilitate future research -- the user assumes all liability.