assist_sim

assist_sim combines musculoskeletal (MSK) models with assistive devices for MuJoCo simulation. It is the middle layer between myo_sim, which supplies the MSK models, and downstream training frameworks such as myoassist.

assist_sim takes a baseline MSK model and a device that YAML describes. It applies the prosthetic surgery (body removals, tendon edits, mesh swaps). Then it attaches the device. It returns a compiled MjModel that you can simulate. assist_sim keeps the baseline MSK model on disk unchanged.

Quickstart

from assist_sim import load_combined
model, data = load_combined("myolegs26", "DephyExoBoot_L1")
# `model` and `data` are ready for mj.mj_step / mj.viewer

This composes the MSK model on demand through myo_sim, which you must install first. It reads the device config from the bundled assist_sim/models/ directory. For full control of the paths, if you already have a baseline MSK XML file on disk:

from assist_sim import load_combined_model

model, data = load_combined_model(
    human_xml="path/to/myolegs26.xml",
    device_config="path/to/your/Device/L1config.yaml",
)

Baseline MSK model, no device

load_msk is the counterpart to load_combined for a model with no device. Use it when you give a bare MSK model to a downstream consumer:

from assist_sim import load_msk

model, data = load_msk("myolegs26")
model, data = load_msk("myolegs26", export_xml="myolegs26.xml")

It skips the device half of the pipeline (surgery, attachment, actuators, tendons, equalities, contacts, sensors). It goes directly from the resolved spec to a compile. Without surgery, the qpos and DOF (degree of freedom) layout stays the same. Therefore the keyframes also need no decomposition and no rebuild. It takes the same optional cache_dir= argument as load_combined.

Or from the CLI:

python -m assist_sim combine myolegs26 DephyExoBoot_L1 -o combined.xml
python -m assist_sim msk myolegs26 -o out.xml      # baseline MSK, no device
python -m assist_sim list
python -m assist_sim validate myolegs26 DephyExoBoot_L1

To inspect any combination visually:

python examples/quickstart.py myolegs26 DephyExoBoot_L1

Available Combinations

The MSK keys mirror the myo_sim model names. All four keys are connected and tested: myolegs22 (planar, 22 muscles, a runtime 26→22 reduction), myolegs26 (26 muscles, passive torso), myolegs (80 muscles, passive torso) and myofullbody (full body). Every device works with every MSK model, because all four MSK models carry the passive torso scaffold. Each pair also has a frozen smoke signature.

See docs/available-models.md for a description of each device and the tested combinations.

Collaboration environments (upper-body)

assist_sim also ships upper-body collaboration environments next to the modular lower-limb devices. Dedicated functions in assist_sim/upper_body.py build them, not load_combined. These environments are not registry devices, and they are not modular:

from assist_sim.upper_body import (
    build_wheelchair,
    build_mpl,
    build_auxivo_liftsuit,
    build_bionic_bimanual,
)

model, data = build_wheelchair(arms="both", torso="passive")  # "both"|"right"|"left"; "passive"|"muscled"
model, data = build_mpl()               # standalone bimanual MPL robot (no human)
model, data = build_auxivo_liftsuit()   # passive back-exosuit on the muscled myotorso
model, data = build_bionic_bimanual()   # MyoChallenge arm + MPL-prosthesis manipulation task

All four environments are available today. The Wheelchair is a seated human who propels a manual wheelchair. The MPL is a standalone bimanual Modular Prosthetic Limb robot. The AuxivoLiftsuit is a passive back-exosuit on the muscled myotorso. bionic-bimanual is the MyoChallenge manipulation task that pairs a biological arm with an MPL prosthesis.

Each builder returns (MjModel, MjData). The three composed environments also give a build_*_spec() companion that returns the uncompiled MjSpec. export_upper_body_xml(spec, path) serializes that spec to a standalone XML file that you can reload. See docs/collaboration-environments.md.

Installation

# Clone
git clone https://github.com/neumovelab/assist_sim.git
cd assist_sim

# Editable install + myo_sim dependency
pip install -e .
pip install git+https://github.com/MyoHub/myo_sim.git   # interim, until on PyPI

Requirements: Python ≥ 3.10, MuJoCo ≥ 3.3.4 (for in-memory MjSpec.delete).

For development:

pip install -r requirements-dev.txt
pytest

Documentation

License

Apache 2.0. See LICENSE.

Citation

If you use this in academic work, cite the parent project myoassist (citation TBD).