MuscleDrivenArm — Classical Controllers

Physics-based control of a planar two-link, six-muscle human arm. This is the reference implementation behind the paper

Physics-Based Control of a Planar Two-Link, Six-Muscle Human Arm: Classical Controllers with a Common Redundancy Layer.

It couples a full musculoskeletal model — arm anatomy, Hill-type muscle mechanics, tendon-excursion kinematics, Euler–Lagrange dynamics, and identified parameters — to four classical controllers through one shared redundancy-resolution layer. Each controller outputs a desired joint torque; a single muscle-force allocation problem then maps that torque to admissible muscle forces via the posture-dependent moment-arm matrix W(θ).

🌿 Two branches, two controller families

Branch	Contents
main (this branch)	Classical / physics-based controllers: impedance (PD+IF), passivity-based, sliding-mode, operational-space, plus the benchmark, sensitivity and EMG-validation pipelines. Lightweight — no bundled data.
learning	Learning-based controllers (MotorNet, muscle synergies, ANFIS, behaviour-cloning, MPC/RL) with the trained models and datasets bundled in.

| Both branches share the same plant (lib/, model_lib/, muscles/) and simulation infrastructure, so results are directly comparable. Switch with git checkout learning.

Controllers on this branch

Controller	Module	Paper family
Impedance / PD+IF	controller/numpy/pd_if_controller.py	Impedance control
Passivity-based (energy tank)	controller/numpy/energy_tank_controller.py	Passivity-based control
Sliding-mode	controller/numpy/sliding_mode.py	Sliding-mode control
Operational-space	controller/numpy/osc_controller.py	Operational-space control
Predictive (benchmark lead)	controller/numpy/predictive.py	—

Each has a NumPy reference implementation and, where relevant, a Torch counterpart under controller/torch/.

Layout

controller/        classical controllers (numpy + torch)
lib/  model_lib/  muscles/   shared plant: skeleton, Hill muscles, dynamics
sim/  trajectory/  tasks/     simulation loop, min-jerk & Lissajous refs, reach tasks
utils/  logging_tools/  plotting/   helpers, run logging, figures
config.py                     plant / gain / numerics configuration
scripts/
  PD_IF/  PASSIVITY/  SLIDING/  OSC/   per-controller reaching runs
  BENCHMARK/                          cross-controller benchmark + robustness
sensitivity/                          Section-VIII Sobol/Monte-Carlo parameter study
figure_repro_emg/  emg_*.py           Section-XI EMG-validation figures
kinarm_replay_validation.py           closed-loop replay of measured KINARM reaching
tests/                                unit tests for the plant and controllers
docs/                                 audit, coverage ledger, optimization notes

Quick start

python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

# a single reaching run with the impedance (PD+IF) controller
python scripts/PD_IF/main_random_reach.py

# the full cross-controller benchmark
python scripts/BENCHMARK/run_benchmark.py

# Section-VIII parameter sensitivity study (Sobol / Monte-Carlo)
python -m sensitivity.sobol_analysis
python -m sensitivity.mc_propagation

Citation

If you use this code, please cite the paper (BibTeX to be added on publication).

License

Released under the MIT License.