minimalist_compliance_control

Project Page | Paper | Tweet/X

A lightweight package for MuJoCo-based compliance control and wrench estimation.

The method estimates external wrenches from motor current/voltage and Jacobians, requires no force sensors or learning, and is plug-and-play with VLM, imitation, and model-based policies across tasks like wiping, drawing, scooping, and in-hand manipulation.

Overview

minimalist_compliance_control provides:

• Wrench simulation and Jacobian utilities,
• Online wrench estimation,
• Compliance reference integration,
• Unified policy/controller orchestration.

Installation

conda create -n mcc python=3.10
conda activate mcc
pip install -e .

For policy stacks (model-based / diffusion-policy / VLM):

pip install -e ".[policy]"

To include the Dynamixel C++ backend:

pip install -e ".[policy]" --config-settings=cmake.define.BUILD_DYNAMIXEL=ON

Policy Scripts

policy/run_policy.py is the main entrypoint for all policy variants. For best performance, make sure your policy loop can sustain 50 Hz. This repository is designed primarily for real-world execution, and real-world performance is often better than in simulation.

Compliance

• policy/compliance.py: base compliance policy implementation (--policy compliance).
• Examples:

  mcc-run-policy --policy compliance --robot leap --sim mujoco --vis view
  mcc-run-policy --policy compliance_model_based --robot toddlerbot --sim mujoco --vis view
  mcc-run-policy --policy compliance_dp --robot toddlerbot --sim real --ckpt /path/to/ckpt.pth
  mcc-run-policy --policy compliance_vlm --robot toddlerbot --sim real --object "star" --site-names "right_hand_center"
  mcc-run-policy --policy compliance --robot toddlerbot --sim real --vis none
  mcc-run-policy --policy compliance --robot arx --sim real --vis none
  mcc-run-policy --policy compliance --robot g1 --sim real --vis none --ip en0

  • Important: For best performance, make sure your policy loop can sustain 50 Hz.
  • Equivalent direct script invocation:

  python policy/run_policy.py --policy compliance --robot leap --sim mujoco --vis view

Real-World Setup

For ARX, install the external ARX5 SDK: https://github.com/real-stanford/arx5-sdk. If you cannot find the CAN interface, try sudo ip link set up can0 type can bitrate 1000000. Example ARX real-run environment:

export ARX5_SDK_PATH=/home/haochen/arx5-sdk/python
export AMENT_PREFIX_PATH=/home/haochen/miniforge3/envs/arx-py310
export ARX5_INTERFACE=can0   # optional but recommended
mcc-run-policy --policy compliance --robot arx --sim real

For ToddlerBot, LEAP, and Unitree G1, real-world support is included in this repository.

Compliance With VLM (Real World Only)

• policy/compliance_vlm.py
  • VLM-guided compliance implementation (--policy compliance_vlm).
• policy/run_affordance_prediction.py
  • Offline affordance prediction + EE pose planning from stereo images in assets/.
• Examples:

  python policy/run_affordance_prediction.py --robot toddlerbot --task wipe
  python policy/run_affordance_prediction.py --robot leap --task draw --site rf_tip if_tip --object "star"
  mcc-run-policy --policy compliance_vlm --robot toddlerbot --sim real

Compliance With DP (Real World Only)

• policy/compliance_dp.py
  • Diffusion-policy compliance implementation (--policy compliance_dp).
• Example:

  mcc-run-policy --policy compliance_dp --robot toddlerbot --sim real --ckpt /path/to/ckpt.pth

Compliance With Model-Based Planning

• policy/compliance_model_based.py
  • Model-based policy selector wrapper (--policy compliance_model_based).
• policy/compliance_model_based_leap.py
  • LEAP-specific model-based implementation.
• policy/compliance_model_based_toddlerbot.py
  • Toddlerbot-specific model-based implementation.
• Example:

  mcc-run-policy --policy compliance_model_based --robot toddlerbot --sim mujoco --vis view

Data and Checkpoints

• Download shared assets from Google Drive.
• Extract robologger.zip into datasets/, then train with:

  python diffusion_policy/train.py --dataset-paths datasets/robologger/compliance_follower_20260110_140133

• Extract toddlerbot_2xm_dp_20260113_141759.zip into results/ for the diffusion policy checkpoint.
• Place the foundation stereo engine at ckpts/foundation_stereo_vitl_480x640_20.engine.
• Set API keys for affordance/compliance providers: GOOGLE_API_KEY and OPENAI_API_KEY (if using the OpenAI provider).

Related Folders

• policy/: policy implementations and policy utilities.
• sim/: simulation adapters (base_sim.py, sim.py) used by run_policy.py.
• hybrid_servo/: model-based algorithms and utilities.
• diffusion_policy/: diffusion model components.
• vlm/: VLM affordance/depth/servers.
• real_world/: hardware adapters (real_world_dynamixel.py, real_world_arx.py, real_world_g1.py) and IMU/camera interfaces.

Related Projects

• ToddlerBot
• Robot Trains Robot
• Locomotion Beyond Feet

Citation

@misc{shi2026minimalist,
  title = {Minimalist {{Compliance Control}}},
  author = {Shi, Haochen and Hu, Songbo and Hou, Yifan and Wang, Weizhuo and Liu, Karen and Song, Shuran},
  year = 2026,
  month = mar,
  number = {arXiv:2603.00913},
  eprint = {2603.00913},
  primaryclass = {cs},
  publisher = {arXiv},
  doi = {10.48550/arXiv.2603.00913},
  urldate = {2026-03-03},
  archiveprefix = {arXiv},
  keywords = {Computer Science - Robotics}
}