Latent Action World Models for Control with Unlabeled Trajectories
TLDR
Latent-action world models learn from both labeled and unlabeled trajectories, enabling efficient control with few action labels.
Reasoning
The paper introduces a novel latent-action representation to align action-conditioned and action-free data, showing strong results on simulated benchmarks. However, it lacks real-world validation and does not address scalability to complex environments.
Read-first score
Read-first score 55.9, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 44.
Field roles
Rank sensitivity
Stability: volatile; rank range: 359.
Keyword Scores
Deep Analysis
Innovations
- Joint use of action-conditioned and action-free data via a shared latent action representation
- Latent-action policy learned through offline RL, bridging offline RL and action-free training
- Enables training on both passive and interactive data for more efficient world model learning
Methodology
The paper introduces a family of latent-action world models that learn a shared latent action representation from both action-conditioned and action-free data. This latent space aligns observed control signals with actions inferred from passive observations, allowing a single dynamics model to train on large-scale unlabeled trajectories with only a small set of action-labeled ones. The latent-action world model is then used to learn a latent-action policy via offline reinforcement learning.
Key Results
On the DeepMind Control Suite, the approach achieves strong performance while using about an order of magnitude fewer action-labeled samples than purely action-conditioned baselines.