Simulation Distillation: Pretraining World Models in Simulation for Rapid Real-World Adaptation
TLDR
SimDist pretrains world models in simulation and adapts them to real-world robot tasks via supervised system identification, enabling rapid improvement.
Reasoning
The paper presents a novel framework that leverages physics simulators for scalable pretraining of world models, addressing data scarcity in real-world robot learning. Its strengths include clear methodology and strong empirical results on contact-rich tasks, but it may be limited by reliance on simulator fidelity and task-specific assumptions.
Read-first score
Read-first score 49.1, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 51.
Field roles
Rank sensitivity
Stability: volatile; rank range: 375.
Keyword Scores
Deep Analysis
Innovations
- Simulation Distillation (SimDist) framework that uses physics simulators as a scalable source of action-conditioned robot experience for pretraining world models
- Transfer of encoder, reward model, and value function from simulation to real world, updating only the latent dynamics model during real-world adaptation
- Reduction of real-world adaptation to supervised system identification while preserving dense, long-horizon planning signals
Methodology
SimDist pretrains a world model in simulation by distilling structural priors from a physics simulator, enabling planning from raw real-world observations. During real-world adaptation, the encoder, reward model, and value function are transferred from simulation, and only the latent dynamics model is updated using real-world prediction losses.
Key Results
Across contact-rich manipulation and quadruped locomotion tasks, SimDist rapidly improves with experience, while prior adaptation methods struggle to make progress or degrade during online finetuning.