arXiv News

First-hand information for everyone

Site

AboutPrivacy PolicyTerms of Use

© 2026 arXiv News

RoC Boosts Robot Task Success by 24% on RMBench Using Selective Agent Intervention | arXiv News

arXiv News

EnglishJapanese

Switch language

EnglishJapanese
Loading account…

First-hand information for everyone

Back to front page
RoboticsEnglishPublished

RoC Boosts Robot Task Success by 24% on RMBench Using Selective Agent Intervention

November 21, 2026arXiv: 2611.26341v1

RoC (Reasoning on Call) is a new framework that lets a robot call in a powerful reasoning agent only when needed. This selective call saves computation and speeds up execution.

The core idea is a three‑part system. A low‑level policy runs most of the time. An online task‑state monitor watches the robot’s progress. When the monitor sees a problem, it asks the reasoning agent to decide whether the robot should redo a move, change its plan, or do both.

On the RBI benchmark, RoC achieved a 99.6% task success rate, a 24‑percentage‑point lift over the best agent‑based baseline. Compared with that baseline, RoC also cut token usage by 65.3%. In addition, against the strongest policy‑based system, RoC improved success rate by 15.1 points.

The improvement comes mainly from efficient use of the reasoning agent. In cases that needed the agent, RoC reconstructed execution or revised plans without re‑calling the agent for every step.

However, real‑robot tests showed that failures still occur when the robot cannot grasp objects, even after the agent’s input. These difficult grasps remain a bottleneck, highlighting the need for stronger visual grasp reasoning.

In short, RoC demonstrates that limited, well‑timed agent calls can dramatically raise success rates while keeping computational costs low. Future work will focus on improving grasp robustness to overcome the remaining failure modes.

Source paper

Reasoning on Call: Efficient Agentic Robot Control for Long-Horizon Robot Manipulation

Robotics
View on arXiv