LIBERO Robot Manipulation Chain-of-Thought Annotations

LIBERO-R Datasets contain chain-of-thought textual reasoning annotations for the LIBERO-100 robot manipulation benchmark. NVIDIA researchers created these annotations to align vision-language-action models via runtime reasoning-action verification. The dataset page was last updated in January 2026.

Use Cases

• Training models to generate step-by-step reasoning annotations for robot manipulation sequences.
• Verifying runtime alignment between textual reasoning chains and robot action execution in the LIBERO-100 benchmark.
• Analyzing the correlation between specific annotated reasoning steps and successful task completion in manipulation episodes.

Strengths

• Annotations are designed for the established LIBERO-100 benchmark, comprising 100 manipulation tasks.
• Provides step-by-step chain-of-thought reasoning aligned with robot action sequences.

Limitations

• The specific number of annotated episodes, rows, or columns is unknown.
• The dataset's utility is limited to users familiar with the LIBERO-100 benchmark environment.

Provenance

Source

NVIDIA research team, as cited in the associated arXiv paper.

Collection Method

Annotations were created for the LIBERO-100 benchmark, as described in the paper 'Do What You Say: Steering Vision-Language-Action Models via Runtime Reasoning-Action Alignment Verification'.

Freshness

Dataset page was updated January 14, 2026.