ManiSkill2: Robotic Manipulation Benchmark with 4M+ Demonstration Frames

4M+ demonstration frames across 20 task families form a unified benchmark for learning generalizable robotic manipulation skills. The dataset, powered by the SAPIEN physics simulator, features over 2000 diverse object models. It was created by haosulab and last updated on March 9, 2025.

Use Cases

• Training imitation learning models based on the 4M+ demonstration frames.
• Benchmarking robotic manipulation algorithms across the 20 out-of-box task families.
• Developing generalizable skills for robots using the diverse object models.
• Evaluating fast visual input learning algorithms as mentioned in the description.

Strengths

• Contains over 4 million demonstration frames.
• Covers 20 distinct task families for robotic manipulation.
• Includes more than 2000 diverse object models.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• The description notes that ManiSkill 3 uses different data, suggesting this version may be superseded.

Provenance

Source

haosulab

Collection Method

Likely generated within the SAPIEN physics simulation environment.

Freshness

Last updated 2025-03-09 09:18:25; freshness should be verified.