Disentangled Robot Learning via Separate Forward and Inverse Dynamics Pretraining

arXiv 2026 39 method, application

TLDR

DeFI decouples forward and inverse dynamics pretraining for robot learning, using video prediction and latent action inference, achieving state-of-the-art results on benchmarks and real-world tasks.

Reasoning

The paper presents a novel framework that separates visual forward dynamics (video prediction) and inverse dynamics (action inference) to leverage action-free videos, showing strong empirical results. However, the abstract does not detail limitations or comparisons to other world model approaches, and the term 'world model' is not explicitly used.

Read-first score

Read-first score 39, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 41.

Recency 6%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Methodology quality 18%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=benchmark,experiment

Topical relevance 29%

58.6

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Reproducibility 18%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=none

Citation impact 18%

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. cited_by_count=0

Citation velocity 12%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 245.

Keyword Scores

world dynamics prediction

video world model

world model

generative world model

world simulator

model-based reinforcement learning world model

interactive world model

Deep Analysis

Innovations

Disentanglement of forward (video generation) and inverse (action prediction) dynamics pretraining to leverage action-free web videos
General Forward Dynamics Model (GFDM) pretrained on diverse human and robot videos for future prediction
General Inverse Dynamics Model (GIDM) trained via self-supervised learning to infer latent actions from unlabeled video transitions
Integration of GFDM and GIDM into a unified architecture for end-to-end fine-tuning on downstream tasks

Methodology

DeFI pretrains two separate models: GFDM for forward dynamics (video generation) on human and robot videos, and GIDM for inverse dynamics (latent action inference) using self-supervised learning on unlabeled video transitions. The pretrained models are then integrated into a unified architecture and fine-tuned end-to-end on downstream robot tasks.

Key Results

DeFI achieves state-of-the-art performance with an average task length of 4.51 on CALVIN ABC-D, 51.2% success rate on SimplerEnv-Fractal, and 81.3% success rate in real-world deployment, significantly outperforming prior methods.

Resources

Paper