RoboShackles: 1,200 Safety-Critical Robotic Video Clips for Testing Embodied AI

RoboShackles is a safety benchmark for evaluating Embodied Foundation Models. The public test split contains 1,200 safety-critical robotic video clips, with 200 videos per category. The dataset was created by YZW00 and last updated on Hugging Face in June 2026.

Use Cases

• Benchmarking model refusal of unsafe robot instructions based on video scenarios.
• Evaluating risk anticipation for human injury before action execution.
• Training or fine-tuning safety classifiers for robotic control systems.

Strengths

• Contains 1,200 video clips for testing.
• Structured with 200 videos per safety-critical category.
• Designed for a specific, high-stakes evaluation task in embodied AI.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Only the test split is publicly released; the full 10,000-clip dataset is not available.

Provenance

Source

YZW00 via Hugging Face.

Collection Method

Likely contains curated or simulated safety-critical robotic scenarios.

Freshness

Last updated 2026-06-22 06:04:05; freshness should be verified.