Sensorimotor World Models: Perception for Action via Inverse Dynamics

arXiv 2026 53 method

TLDR

Introduces sensorimotor world model with inverse dynamics regularization for action-aligned latent representations and stable planning.

Reasoning

Strengths include a simple yet effective regularization method that prevents representation collapse and aligns latent states with actions, demonstrated on 2D/3D tasks. Weaknesses are limited to simple simulated environments without real-world validation or comparison to more complex baselines.

Read-first score

Read-first score 53, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 35.

Recency 6%

100

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

Citation impact 18%

95.4

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.95447039

Topical relevance 29%

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

Methodology quality 18%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=none

Reproducibility 18%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=code

Citation velocity 12%

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

Field roles

FoundationFrontierBridge

Rank sensitivity

Stability: volatile; rank range: 437.

Keyword Scores

world model

model-based reinforcement learning world model

world dynamics prediction

generative world model

interactive world model

world simulator

video world model

Deep Analysis

Innovations

Inverse dynamics regularization for latent world models to prevent representation collapse and induce action-aligned representations
Training stable latent world models from offline, reward-free trajectories without frozen encoders, exponential moving averages, or complex latent regularizers
Biasing latent representations toward controllable degrees of freedom while discarding uncontrollable distractors

Methodology

The paper proposes a sensorimotor world model (SMWM), a latent JEPA-style world model trained end-to-end with inverse dynamics regularization. The model is trained on offline, reward-free trajectories, using the inverse dynamics loss to force latent states to preserve information about the action underlying a transition, thereby preventing collapse and aligning representations with action-relevant features. No frozen encoders, exponential moving averages, or complex latent regularizers are used.

Key Results

SMWM learns compact, interpretable latent spaces and enables competitive planning performance across simple 2D and 3D control tasks.

Limitations

Only evaluated on simple 2D and 3D control tasks, leaving scalability to more complex environments unaddressed
Potential limitations of inverse dynamics regularization in capturing all relevant features for action in highly stochastic or partially observable settings

Resources

Paper