Modified LIBERO RLDS: Four Robotics Datasets for OpenVLA Fine-Tuning

Four modified LIBERO robotics datasets formatted in RLDS for OpenVLA fine-tuning experiments, released by the OpenVLA team in 2024. These datasets adapt the original LIBERO benchmark specifically for training and evaluating vision-language-action models in simulated environments.

Use Cases

• Fine-tuning OpenVLA models using RLDS-formatted trajectories
• Benchmarking robotic manipulation performance on modified LIBERO tasks
• Training vision-language-action policies using language instructions

Strengths

• Standardized RLDS format
• MIT licensed for open research
• Directly supports Arxiv 2406.09246 reproduction

Limitations

• Restricted to simulated LIBERO environments
• Modifications from original LIBERO benchmark may limit direct comparisons
• Metadata lacks explicit row counts and column definitions

Provenance

Source

OpenVLA Research (Arxiv 2406.09246)

Collection Method

Simulated environment data generation

Freshness

Last updated September 2024; linked to the 2024 OpenVLA research publication.