Reproduce “Generating Freeform Endoskeletal Robots” on OpenResearch

About this paper

Muhan Li, Lingji Kong, Sam Kriegman

Researchers at Northwestern University developed a computational platform that automatically designs and co-optimizes endoskeletal robots by integrating rigid internal skeletons with deformable soft tissues. This system uses a novel multi-physics simulator, a deep variational autoencoder for morphology, and a universal reinforcement learning controller to generate diverse, functional robots capable of complex terrestrial locomotion.

Key insights

• The VAE's learned latent space is smooth and expressive, allowing for efficient evolutionary search and revealing 'tunable genes' that correlate with macroscopic morphological traits.
• A single universal reinforcement learning controller can effectively generalize and adapt to an evolving population of robots with vastly different geometries and topologies, improving computational efficiency.
• The combined evolutionary optimization of both morphology and control is crucial for achieving high-performing endoskeletal robot designs, significantly outperforming random morphological search.

Summary by alphaXiv.