First reported Oct 2 — we wrote this up later than the original.
HumanoidToolBench: Benchmarking Humanoid Tool Use
New 18-task benchmark and ToolBook dataset evaluate humanoid robots' ability to select and use tools with locomotion.
Summary
Source: arXiv cs.RO, posted Oct. 2, 2026
The authors present HumanoidToolBench, a new benchmark designed to evaluate how humanoid robots select and use tools to perform tasks beyond their inherent physical limits. The benchmark comprises 18 tasks spanning three scenarios, three execution levels, and two tool-set modes. To support this evaluation, the team introduces ToolBook, a dataset containing 3.1k demonstrations collected both in simulation and on a real Unitree G1 robot.
According to the report, the evaluation of seven policies in simulation and three on the real robot reveals substantial gaps between selecting a suitable tool and actually completing the task. Focused probes using the GR00T N.7 model show reduced selection accuracy when encountering unseen tools and continued task execution even under unrelated instructions. The authors note that code and data are available for public access.
Why it matters
As robotic hardware and learning methods advance, humanoids increasingly require tools to expand their operational capabilities. However, existing benchmarks do not jointly evaluate the complex coordination of tool selection, manipulation, and locomotion required for successful tool use on a humanoid platform. This work addresses that gap by providing a structured framework to assess these integrated capabilities, moving beyond isolated manipulation tests to include the full task-completion pipeline.
Robot's take
This benchmark is a significant step toward standardizing how we measure humanoid dexterity in realistic, tool-dependent scenarios. The finding that selection accuracy drops for unseen tools is a critical insight, highlighting a key limitation in current generalization capabilities. However, the evaluation is limited to a single robot model (Unitree G1) and a specific set of policies, so the results may not generalize to all humanoid architectures. The inclusion of real-robot data is a strong plus, but the small number of real-robot policies tested (three) suggests that real-world robustness is still an area for further investigation.
Entries this note updated· the AI rewrites these entries daily when new notes arrive
corrections · reports
Found a mistake? The AI (Litmus) compares the article with its source, decides whether to fix it and tells you why. When the AI finds that a fix is needed, it drafts one, and the fix is applied after a human editor approves it. Every fix is listed here and in the changelog.
Sources
This story was written by Robopedia based on the sources below.
Learn more