Compact Visuotactile World Models for Lifting: Prediction, Reward Alignment, and Force Constraints

arXiv 2026 39.1 method, application

TLDR

Compact visuotactile world models improve force-constrained lifting control via model-assisted feedback, but not imagined RL; evaluated extensively in MuJoCo without physical robots.

Reasoning

Strengths include systematic evaluation across many policies and environments, confidence intervals, and a compact model. Weaknesses are the lack of physical-robot transfer, no closed-loop safety guarantee, and the finding that imagined RL underperforms reactive IQL, limiting practical impact.

Read-first score

Read-first score 39.1, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 51.

Recency 5%

100

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

Topical relevance 25%

72.9

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 15%

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

Reproducibility 15%

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

Graph prestige 15%

5.7

PageRank-style prestige over the local citation graph built from available OpenAlex references. local_pagerank=0.0016

Citation impact 15%

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

Citation velocity 10%

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

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 485.

Keyword Scores

world model

world dynamics prediction

model-based reinforcement learning world model

interactive world model

world simulator

generative world model

video world model

Deep Analysis

Innovations

Compact 652,157-parameter action-conditioned visuotactile world model for rigid-box lifting under force constraints.
Model-assisted force feedback using learned visuotactile dynamics to improve force-budgeted success.
Fixed evaluation protocol spanning 34 policies, 120 fresh MuJoCo environments with geometry/physical-parameter shifts, and 324 independently replayed action branches on 12 ID environments.
Empirical tactile-residual stress test with 330 additional executions.
Evidence that accurate tactile forecasts need not improve pooled force-constrained control.

Methodology

The study uses a 652,157-parameter action-conditioned visuotactile world model with matched behavior cloning, policy learning in imagination, independent reactive implicit Q-learning, and model-assisted force feedback. A fixed protocol executes 34 policies on 120 fresh MuJoCo environments with geometry and physical-parameter shifts, plus 324 independently replayed action branches on 12 additional in-distribution environments. Evaluation metrics include force action-effect MAE, force-budgeted success, and pooled joint success, with persistence and reactive IQL as comparison points.

Key Results

Visuotactile dynamics reduce force action-effect MAE from 0.413 N for persistence to 0.338 N. Model-assisted feedback raises in-distribution force-budgeted success from 73.3% to 93.3% (paired difference +20.0 pp, 95% CI [6.7, 33.4]) mainly during scripted lowering, while the pooled difference is +3.9 pp (95% CI [-4.5, 11.7]); imagined RL achieves 11.9% pooled joint success versus 25.0% for reactive IQL.

Limitations

Evidence is limited to rigid-box lifting after a common approach.
No physical-robot transfer is reported.
No closed-loop safety guarantee is provided.
The pooled benefit of model-assisted feedback is not statistically significant (95% CI includes zero).
Imagined RL underperforms reactive IQL (11.9% vs 25.0% pooled joint success).

Resources

Paper