morphology-adaptive muscle-driven locomotion via attention mechanisms

This repository demonstrates how a single locomotion controller can work with different shapes (biped and quadruped). It uses an attention mechanism to handle an arbitrary number of inputs, and the same module is shared across all muscles to support an arbitrary number of outputs.

More details in this paper.

View animation in higher quality here.

For simulation, this repository uses Algovivo, originally built for the browser using WebAssembly, but here a native build is used to enable PyTorch integration.

run on GitHub Actions

The workflow trajectory-attn.yml runs the controller on both morphologies and generates a video. If you have your own copy or fork of this repository, you can run the workflow from the GitHub Actions UI, no local installation needed. Once it completes, the video will be saved as a workflow artifact and can be downloaded from the workflow run page.

run locally

Pull the Docker image:

docker pull ghcr.io/juniorrojas/morphology-adaptive/bundle:95853e0

Generate trajectory (use biped or quadruped):

docker run --rm \
  --user $(id -u):$(id -g) \
  -v $(pwd):/workspace \
  -w /workspace \
  ghcr.io/juniorrojas/morphology-adaptive/bundle:95853e0 \
  python /morphology-adaptive/scripts/generate_trajectory_with_attn_policy.py \
  --agent /morphology-adaptive/data/agents/biped \
  --policy /morphology-adaptive/data/policies/attn \
  --steps 200 \
  -o trajectory.out

Render frames:

docker run --rm \
  --user $(id -u):$(id -g) \
  -e HOME=/tmp \
  -v $(pwd):/workspace \
  -w /workspace \
  ghcr.io/juniorrojas/morphology-adaptive/bundle:95853e0 \
  node /morphology-adaptive/algovivo.repo/utils/trajectory/renderTrajectory.js \
  --mesh ./trajectory.out/mesh.json \
  --steps ./trajectory.out/steps \
  --width 300 \
  --height 300 \
  -o frames.out

Make video:

ffmpeg -y \
  -framerate 30 \
  -i frames.out/%d.png \
  -c:v libx264 \
  -profile:v high \
  -crf 20 \
  -pix_fmt yuv420p \
  video.out.mp4

citation

@inproceedings{10.1145/3712255.3734277,
  author = {Rojas, Junior},
  title = {Morphology-Adaptive Muscle-Driven Locomotion via Attention Mechanisms},
  year = {2025},
  isbn = {9798400714641},
  publisher = {Association for Computing Machinery},
  address = {New York, NY, USA},
  url = {https://doi.org/10.1145/3712255.3734277},
  doi = {10.1145/3712255.3734277},
  booktitle = {Proceedings of the Genetic and Evolutionary Computation Conference Companion},
  pages = {2138–2142},
  numpages = {5},
  keywords = {neural networks, attention mechanisms, virtual creatures},
  location = {NH Malaga Hotel, Malaga, Spain},
  series = {GECCO '25 Companion}
}