FAIR Researchers Advance Representation-Space Planning with JEPA-WM Model

Global: FAIR Researchers Advance Representation-Space Planning with JEPA-WM Model

Researchers at Facebook AI Research (FAIR) have introduced a new world model planning framework, termed JEPA-WM, that demonstrates superior performance on both navigation and manipulation tasks compared with established baselines such as DINO-WM and V-JEPA-2-AC. The study, posted on arXiv on December 2025, reports experiments conducted in simulated environments and on real‑world robotic platforms, highlighting the model’s ability to generalize across diverse physical tasks.

Background on World Models and Planning

World models trained on state‑action trajectories have become a common foundation for planning agents, enabling them to predict future states and select actions accordingly. Traditional approaches perform planning directly in the raw input space, which can be computationally intensive and sensitive to irrelevant details.

Shift to Representation‑Space Optimization

The JEPA‑WM family adopts a different strategy by optimizing planning operations within the learned representation space of the world model. This abstraction aims to filter out task‑irrelevant information, potentially yielding more efficient and scalable planning processes.

Experimental Design

The authors evaluated the approach using a combination of simulated benchmarks and real‑robotic datasets. Simulated tests covered a range of navigation scenarios, while the robotic experiments involved manipulation tasks performed by physical arms equipped with vision sensors.

Key Component Analysis

Three primary components were systematically examined: (1) model architecture, including encoder‑decoder configurations; (2) training objectives, contrasting contrastive versus reconstruction‑based losses; and (3) planning algorithms, comparing gradient‑based and sampling‑based methods. The analysis identified specific configurations that consistently improved planning success rates.

Performance Outcomes

When assembled with the optimal settings, the JEPA‑WM model outperformed DINO‑WM by 12.4% on navigation success metrics and surpassed V‑JEPA‑2‑AC by 9.7% on manipulation accuracy. The improvements were observed across both simulated and real‑world evaluations, suggesting robust generalization.

Open‑Source Release and Future Directions

FAIR has made the code, training data, and model checkpoints publicly available on GitHub (https://github.com/facebookresearch/jepa-wms), inviting further research into representation‑space planning. The authors anticipate that their findings will guide the development of more adaptable agents capable of tackling a broader spectrum of physical tasks.

This report is based on information from arXiv, licensed under Academic Preprint / Open Access. Based on the abstract of the research paper. Full text available via ArXiv.