ManiSkill: Benchmark for Robotic Manipulation with 4M+ Demonstration Frames

ManiSkill is a unified benchmark for learning generalizable robotic manipulation skills powered by the SAPIEN simulator. It features 20 out-of-box task families with over 2000 diverse object models and more than 4 million demonstration frames. The dataset, created by haosulab and last updated in March 2024, is designed to enable fast visual input learning, supporting CNN-based policies that can collect samples at about 2000 FPS.

Use Cases

• Benchmarking visual reinforcement learning algorithms based on the 20 provided task families.
• Training policies for sim-to-real transfer using the diverse object models and demonstration frames.
• Studying sample efficiency in robotic control based on the high-speed simulation environment.

Strengths

• Includes 20 distinct task families for varied manipulation challenges.
• Contains over 2000 diverse object models for generalization testing.
• Provides more than 4 million demonstration frames for imitation learning.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• The specific file formats and total data size are unknown, which may complicate integration.

Provenance

Source

haosulab

Collection Method

Generated within the SAPIEN simulation environment.

Freshness

Last updated 2024-03-15 23:21:41