# FIRM-WM: State-factorized factual-interventional recurrent modeling for reward-free visual planning

arXiv: https://arxiv.org/abs/2609.22816 · HTML: https://mlsanity.com/paper/2609.22816/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Placed in Embodied and robot world models, one of 11 directions in World models.
- Builds on 18 of its 22 references that are in the map.
- First posted on arXiv 19 Sep 2026.

## Placed in

- World models / [Embodied and robot world models](https://mlsanity.com/wm/embodied-and-robot-world-models/)

## Builds on (most cited first, 18 of 18)

- [DINOv2: Learning Robust Visual Features without Supervision](https://mlsanity.com/paper/2304.07193/) (2023)
- [Learning Latent Dynamics for Planning from Pixels](https://mlsanity.com/paper/1811.04551/) (2018)
- [Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture](https://mlsanity.com/paper/2301.08243/) (2023)
- [TD-MPC2: Scalable, Robust World Models for Continuous Control](https://mlsanity.com/paper/2310.16828/) (2023)
- [DINO-WM: World Models on Pre-trained Visual Features enable Zero-shot Planning](https://mlsanity.com/paper/2411.04983/) (2024)
- [LeWorldModel: Stable End-to-End Joint-Embedding Predictive Architecture from Pixels](https://mlsanity.com/paper/2603.19312/) (2026)
- [Learning from Reward-Free Offline Data: A Case for Planning with Latent Dynamics Models](https://mlsanity.com/paper/2502.14819/) (2025)
- [Hierarchical Planning with Latent World Models](https://mlsanity.com/paper/2604.03208/) (2026)
- [Predictive but Not Plannable: RC-aux for Latent World Models](https://mlsanity.com/paper/2605.07278/) (2026)
- [Sub-JEPA: Subspace Gaussian Regularization for Stable End-to-End World Models](https://mlsanity.com/paper/2605.09241/) (2026)
- [NewtonianVAE: Proportional Control and Goal Identification from Pixels via Physical Latent Spaces](https://mlsanity.com/paper/2006.01959/) (2020)
- [Latent Geometry Beyond Search: Amortizing Planning in World Models](https://mlsanity.com/paper/2605.08732/) (2026)
- [Fast LeWorldModel](https://mlsanity.com/paper/2606.26217/) (2026)
- [On the Identifiability of Controlled World Models](https://mlsanity.com/paper/2607.22430/) (2026)
- [Is Forward Prediction Enough? Physical State Grounding for JEPA World Models](https://mlsanity.com/paper/2608.06799/) (2026)
- [SAGE: Subgoal-Conditioned Action Generation for Latent World Model Planning](https://mlsanity.com/paper/2607.17973/) (2026)
- [PhyLatent: Learning Dynamics-Relevant Representations for JEPA World Models](https://mlsanity.com/paper/2608.05720/) (2026)
- [Latent World Models with Monotone Planning Costs for Image-Goal Navigation](https://mlsanity.com/paper/2608.09073/) (2026)

## Abstract (arXiv)

Reward-free latent world models can learn from offline videos and solve new image--goal tasks by optimizing actions through predicted latent futures. This setting places two demands on the planning state: its coordinates must be comparable with a goal image. Moreover, its dynamics must retain velocity, motion trend, contact, and other history--dependent information beyond those goal coordinates. Offline training creates a second mismatch: each recorded trajectory reveals one factual future, whereas a sampling--based planner compares many actions that were not taken from the same state. We introduce FIRM-WM (Factual--Interventional Recurrent World Model), a compact pixel world model designed around these two gaps. Its recurrent state separates a typed, goal--comparable configuration from a 128-dimensional dynamic fiber used for prediction but excluded from the terminal goal cost. Broad factual trajectories provide state coverage, while common--reset intervention branches provide observed outcomes for alternative action sequences. Before executing each branch, we reset the environment and restore the same recorded values exposed by the environment's state--setting interface. Under matched CEM planning and three independent full-pipeline seeds, FIRM-WM reaches 99.0±1.0% on TwoRoom, 92.7±2.1% on Reacher, and 88.0±3.0% on OGBench-Cube, compared with 89.0%, 88.0%, and 70.0% for LeWM. The deployed model uses 2.98--3.42M parameters and records 2.13--11.60× lower planning time on these tasks.