ResearchResearch paperRobotics & Embodied AI1 source · Oct 12, 2026

Embodied Skill Bench: Benchmarking How Well Agents with Embodied Skills Perform on Physical Tasks

LLM agents increasingly control robots using VLA policies, perception models, and procedural knowledge.

On this basis we introduce EmbodiedSkillBench, which runs an agent in a closed observation-action loop and records success, actions, tokens, and wall-clock time.
It supports three modes: native control, skill-only execution, and Agent-Directed Reuse (ADR), a mode we propose in which the agent keeps its native controls and decides for each skill whether to invoke, modify, or ignore it.
Across four agents on 100 perturbed LIBERO-Pro tasks, we find that skills help an agent only as far as they exceed what it can already do.
This suggests that as agents grow stronger, the role of embodied skills shifts from supplying missing capability to reducing cost: GPT-6-Astra already solves most tasks through native control, so restricting it to skills removes more than it adds.

[1]Embodied Skill Bench: Benchmarking How Well Agents with Embodied Skills Perform on Physical Tasks
arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 12, 06:11 AM
LLM agents increasingly control robots using VLA policies, perception models, and procedural knowledge.
On this basis we introduce EmbodiedSkillBench, which runs an agent in a closed observation-action loop and records success, actions, tokens, and wall-clock time.

Extractive summary: sentences quoted from the sources.

Before this