SafeDojo: Safe Reinforcement Learning for VLA via Interactive World Model

arXiv 2026 60.8 method

TLDR

SafeDojo introduces a model-based safe RL framework for VLA policies using an interactive video world model to learn safe actions via imagination.

Reasoning

The paper presents a novel integration of world models with safe RL for vision-language-action policies, demonstrating strong results on the SafeLIBERO benchmark. However, it lacks real-world validation and does not discuss limitations or generalizability beyond the simulated environment.

Read-first score

Read-first score 60.8, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 53.

Recency 6%

100

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

Citation impact 18%

95.2

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.95155193

Topical relevance 29%

75.7

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=baseline,result

Reproducibility 18%

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

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: 491.

Keyword Scores

world model

interactive world model

model-based reinforcement learning world model

video world model

world dynamics prediction

world simulator

generative world model

Deep Analysis

Innovations

First model-based safe reinforcement learning framework for vision-language-action policies using world model-based imagination
Interactive video world model for action-conditioned future predictions
ResNet success classifier to estimate per-step task progress from imagined frames
Lightweight safety head to predict per-step safety costs from latent context and action chunk
Lagrangian-based constrained GRPO objective to balance task reward and safety cost

Methodology

SafeDojo performs online reinforcement learning on top of an interactive video world model that generates action-conditioned future predictions. A tailored ResNet success classifier estimates per-step task progress from imagined frames, and a lightweight safety head predicts per-step safety costs from latent context together with the proposed action chunk. The decoupled task-reward and safety-cost signals are balanced through a Lagrangian-based constrained GRPO objective, enabling coordinated improvement of task success and safety under explicit constraints.

Key Results

On SafeLIBERO, SafeDojo achieves the best aggregate task success, safe success, and execution efficiency among baselines, with an 8.25 percentage-point improvement over the strongest baseline on Level I. Real-world Franka deployment shows the best average task and safe-success rates across five tasks.

Resources

Paper