ROBOSHACKLES: A Safety Dataset for Human-Injury Prevention in Embodied Foundation Models

2026-06-16

AI Summary

Existing embodied foundation models (EFMs) lack real-world safety-aligned data, as collecting robot-induced harm scenarios is ethically constrained. To address this gap, this work proposes a method for constructing safety-critical training data without requiring real hazardous interactions: leveraging real DROID observations, the approach integrates scene understanding, danger-aware image editing, temporal prompt generation, and single-pass video synthesis via Wan2.7 to automatically generate high-fidelity, semantically annotated videos of both direct and indirect harm scenarios. The resulting ROBOSHACKLES dataset comprises 10,000 such clips. Evaluations reveal that all tested EFMs exhibit unsafe behaviors in these safety-critical scenarios, achieving a 100% unsafe rate, thereby underscoring the dataset’s critical role and its value in filling a key void for safe EFM training and evaluation.

Abstract

Embodied Foundation Models (EFMs) integrate multimodal understanding, future-state reasoning, and executable robot actions. Yet their safety alignment for human-injury prevention remains underexplored, primarily because real-world data of robots harming humans or creating hazardous household situations cannot be safely or ethically collected. To address this challenge, we propose a safety-critical data construction pipeline for human-injury prevention in EFMs.Starting from real DROID observations, our construction pipeline proceeds through scene understanding, hazard-aware image editing, temporal prompt generation, and single-pass rollout synthesis. The temporal prompts specify the expected scene evolution, while Wan2.7 synthesizes realistic robotic rollouts from the edited hazardous states in a single pass. Using this pipeline, we construct ROBOSHACKLES, a 10,000-clip robotic video dataset derived from real DROID observations, spanning two direct-harm and four indirect-harm categories. To ensure dataset quality, we assess task completion and visual quality with automatic metrics, and evaluate six representative EFMs under a refusal-based safety criterion. Results show that all evaluated models produce unsafe actions in the tested safety-critical scenarios, yielding a 100% unsafe action generation rate. ROBOSHACKLES serves as a scalable benchmark and training resource for refusal learning and hazard anticipation before robot action execution.The dataset is publicly available at https://huggingface.co/datasets/YZW00/RoboShackles.

Problem

• Embodied Foundation Models
• human-injury prevention
• safety alignment
• hazardous scenarios
• robotic safety

Innovation

• Embodied Foundation Models
• Safety Alignment
• Hazard-Aware Synthesis
• Refusal Learning
• Robotic Safety Dataset