DOM: Dynamic Object Manipulation Benchmark with 200K Episodes

200,000 synthetic episodes across 2,800+ scenes and 206 objects form this large-scale benchmark for robotics. The DOM dataset, created by hzxie, is designed to train and evaluate Vision-Language-Action models on dynamic manipulation tasks. It was last updated on April 27, 2026.

Use Cases

• Training robotic agents for dynamic object manipulation based on the synthetic episodes.
• Evaluating temporal anticipation and rapid perception in robotics based on the benchmark's design.
• Benchmarking Vision-Language-Action models on a large-scale manipulation task based on the 200K episodes.

Strengths

• Large scale with 200,000 episodes for training and evaluation.
• Diverse scene and object coverage with 2,800+ scenes and 206 objects.
• Specifically designed for the challenging domain of dynamic robotic manipulation.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• The data is synthetic, which may limit direct applicability to real-world robotic systems.

Provenance

Source

hzxie on Hugging Face

Collection Method

Synthetically generated.

Time Range

null

Freshness

Last updated 2026-04-27 18:23:54; freshness should be verified.

Geography

null