Reproduce “Motion Primitive Discovery in a Humanoid Robot via Self-Organising Maps for Phase Recognition” 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

Radovan Gregor, Igor Farkaš

Researchers at Comenius University Bratislava developed a biologically inspired, two-level architecture for motor primitive discovery and online phase recognition in a humanoid robot. This system achieved 93.9% phase recognition accuracy using only learned motor primitives and up to 97.3% with full contextual information on a simulated NICO humanoid.

Key insights

  • SOM-discovered motor primitives, when temporally integrated by an ESN, are sufficient to support high online phase recognition accuracy, capturing a dominant share of phase-discriminative information.
  • Arm and hand kinematics encode distinct and complementary aspects of motor behavior, with A-SOM capturing arm transport dynamics and H-SOM effectively encoding hand configurations and grasp types.
  • Contextual signals (e.g., action identity, object distance, contact) refine phase recognition accuracy but play a secondary role compared to the information inherent in primitive sequences, diminishing in value with increased ESN reservoir capacity.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.