RoboTwin Multi: Multi-Embodiment Robot Demonstrations for 19 Tasks

19 distinct robotic manipulation tasks are represented, each with 8 different robot embodiment configurations. The dataset includes 50 demonstrations for each unique combination of task and embodiment, totaling 7,600 demonstrations. Generated by jsw19 and uploaded to Hugging Face, this version excludes the 'put_object_cabinet' task due to stability issues in the final generation run.

Use Cases

• Training cross-embodiment policy transfer models based on the 8 embodiment configurations per task.
• Benchmarking robot learning algorithms on a diverse set of 19 manipulation tasks.
• Studying the effect of robot morphology on task performance using the paired demonstrations across embodiments.
• Developing simulation-to-real transfer methods using the generated demonstration data.

Strengths

• Structured generation from 19 distinct robotic tasks.
• Includes 8 different robot embodiment configurations for each task, enabling comparative analysis.
• Provides 50 demonstrations per task-embodiment pair, offering substantial examples for learning.

Limitations

• Description metadata is limited; actual data quality, format, and content require manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• The 'put_object_cabinet' task is explicitly excluded, indicating potential gaps in task coverage.

Provenance

Source

huggingface

Collection Method

Generated simulation data.

Freshness

Last updated 2026-06-28 10:36:26; freshness should be verified.