Arnold¶

A generalist muscle transformer policy.

Arnold is a transformer policy trained to control musculoskeletal models across 14 manipulation and locomotion tasks spanning four embodiments. This site documents how to install the code, download the released checkpoints, train new policies, evaluate the released ones, and reproduce figures in the paper.

Installation

Docker image or conda environment. Under 30 minutes on a modern machine.
Data and checkpoints

What ships in the repository, and what you need to fetch from Zenodo.
Training

BC, PPO, OBC, OBC-PPO, RL fine-tuning and self-distillation.
Evaluation

Benchmark the released OBC, Arnold and expert policies.

What is included¶

The code and scripts to train policies with BC, PPO, OBC, OBC-PPO, RL fine-tuning and self-distillation — including the expert policies used for imitation learning.
Pretrained checkpoints for every method in the paper, including the ablations, so the results and videos can be reproduced directly.

Model checkpoints and benchmark results¶

All checkpoints, benchmark result files and cached learning curves are hosted on Zenodo:

Zenodo record 21493316

The git repository contains only the code plus small configuration files. Everything else must be unzipped into data/ before running the training, evaluation or plotting scripts — see Data and checkpoints for the exact directory layout.

Reproducing the paper¶

Result	Page
Radar plot, ablation bar plots, and all learning curves	Replicate plots
Effective dimensionality of the learned actions	CSI analysis
Training inside a constrained action subspace	CSI-Finetuning
MT-SAC vs. MT-PPO	Multi-task RL baselines

All figures are written under data/figures/.

Citation¶

If you use Arnold in your research, please cite:

Chiappa, A. S., An, B., Simos, M., Li, C., & Mathis, A. (2025). Arnold: a generalist muscle transformer policy. arXiv:2508.18066. arXiv · PDF

@article{chiappa2025arnold,
  title         = {Arnold: a generalist muscle transformer policy},
  author        = {Chiappa, Alberto Silvio and An, Boshi and Simos, Merkourios and
                   Li, Chengkun and Mathis, Alexander},
  journal       = {arXiv preprint arXiv:2508.18066},
  year          = {2025},
  eprint        = {2508.18066},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  url           = {https://arxiv.org/abs/2508.18066}
}