torch-fabrics

Implementation of Geometric Fabrics using PyTorch with environments in MuJoCo and Gymnasium (formerly OpenAI Gym).

Using this repo

conda env create -f environment.yml
conda activate fabrics

To run an example:

python panda_example.py

The transform tree is based on the same structure as that used in the RMP2 framework.

TODOs

• RK4 integration option during stepping (+accuracy, -speed)
• fabric_solve in fabrics.py uses the naive method (Algorithm 3) from the RMP2 paper. Could we instead improve the complexity by using Algorithm 4?
• Implement more examples of fabrics
• Tune existing fabrics
• Correctness verification (i.e. does this adhere to the design invariants required by fabrics?)
• Create more abstractions for more general purpose fabric design
• swappable transform tree structure

Examples

You'll have to tune fabric parameters because I forgot the parameters I used to get these videos a while ago and the current parameters give different behavior.

Point mass

Panda arm

See more

• Geometric Fabrics for the Acceleration-based Design of Robotic Motion
• Fabrics.jl, a Julia implementation of Geometric Fabrics by Alphonsus Adu-Bredu
• Fabrics (with JAX), a JAX implementation of Geometric Fabrics by Thomas Cohn
• RMP2: A Structured Composable Policy Class for Robot Learning
• RMP2 Codebase