Reproduce “Stable Motion Primitives via Imitation and Contrastive Learning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Rodrigo Pérez-Dattari, Jens Kober
Researchers at Delft University of Technology developed CONDOR, a framework that enables robots to learn globally asymptotically stable motion primitives from demonstrations using arbitrary deep neural networks. By combining multi-step imitation learning with a novel contrastive loss, the system ensures robots reliably reach their goals while maintaining the flexibility and representational power of unconstrained DNN architectures. The method was validated to achieve 100% stability in simulated and real-world robotic tasks across various complexities.
Key insights
- Global asymptotic stability can be enforced in learned robot motions using contrastive learning, without imposing structural constraints on the underlying deep neural network architecture.
- Mapping task-space states to a latent space, where a simple stable system is defined, can implicitly enforce a diffeomorphism and ensure stability in the original space.
- Adaptive gains for the latent stable system significantly improve the stability and generalization capabilities of the learned motion primitives.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.