LetTokenBurn
Sign in
Auto57%Harbor: pendingAgentic / Tool UseMultimodal

RobotWorld

RobotWorld: Benchmarking Multimodal Agents for Robot Use Across Diverse Tasks and Embodiments

A simulation benchmark with 84 tasks across manipulation, locomotion, driving, and aerial control to evaluate multimodal agents' ability to execute physical robot tasks.

Paper

How to run

ltb · robotworld
$ npm i -g lettokenburn-cli
$ ltb bench info robotworld

External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.

Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs

Scores

5 sourced entries
HarnessSourceConfidenceLink
Astra19.0N/AIndependentmediumOct 7, 2026arxiv.org
Claude Opus 5.515.5N/AIndependentmediumOct 7, 2026arxiv.org
Kimi K32.4N/AIndependentmediumOct 7, 2026arxiv.org
DeepSeek V4.1 Flash1.2N/AIndependentmediumOct 7, 2026arxiv.org
Gemini 3.8 Flash1.2N/AIndependentmediumOct 7, 2026arxiv.org

Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.

Reproduced by the community

All runs
Loading community runs…

About

RobotWorld comprises 84 tasks spanning diverse robot embodiments and control demands, featuring explicit interaction budgets and executable success checks. It evaluates agents by linking task outcomes to execution traces, revealing capability gaps in composing perception and control workflows despite local successes.