Learning Intrinsic Dynamics in Latent World Models

Key points

• We investigate the connection between action predictability, defined as the ability to reconstruct an action from consecutive latent states, and the preservation of intrinsic dynamics.1
• Empirically, controlled appearance-diversity comparisons, supervision interventions, and latent probes support the role of action prediction in preserving control-relevant transition information.1
• We instantiate action predictability with Concat IDM, which predicts actions from concatenated latent states and updates the shared encoder using action labels already available in offline trajectories, without changing the test-time planner.1
• Together, these findings support action predictability as an inductive bias toward intrinsic dynamics and demonstrate the practical benefits of Concat IDM for latent planning under visual diversity.1

Sources (1)

• [1]Learning Intrinsic Dynamics in Latent World Models
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 15, 09:57 AM

  Learning intrinsic dynamics---the relationship between actions and their consequences---is essential for world models.

  We investigate the connection between action predictability, defined as the ability to reconstruct an action from consecutive latent states, and the preservation of intrinsic dynamics.

Extractive summary: sentences quoted from the sources.