Native Robot Reasoning: leaderboard

Vision-language models control robots on modified LIBERO tasks and real-world laboratory tasks. Scores separate task success from partial progress and retain each model’s reasoning level. Published general-capability checks of the released base model are separate from its trained robot policy.

Source: arxiv.org.

Benchmark results

Results are published separately for each benchmark or evaluation setting.

Interactive version: theaggregate.ai/benchmark?slug=native-robot-reasoning · How It Works · Data refreshed daily, snapshot 2026-11-17.