How RuoYu Technology Secured China’s First Explosion‑Proof Certification and the World’s First Fueling‑Brain Robot

Robot and Explosion‑proof Certification

RuoYu Technology (Shenzhen) introduced the explosion‑proof special‑purpose robot “RuoYu LanYue 01” powered by its self‑developed “RuoYu Jiutian” robot brain. In a simulated oil‑gas station the robot performed a closed‑loop task chain: voice command reception, environment perception, navigation, equipment‑state recognition and precise manipulation.

In March 2026 the robot obtained the Ex db eb ib mb IIB T4 Gb whole‑machine explosion‑proof certification and the Ex db IIB T5 Gb collaborative‑arm certification, making it the first Chinese wheel‑based humanoid robot with explosion‑proof qualification.

Demonstrated Application Scenarios

Fueling stations : From April 2026 the robot operated at a Foshan gas station, autonomously executing the full fueling workflow (open/close caps, pick up gun, insert, fill, return gun). Each step tolerates only a few millimetres of error and the robot adapts to diverse vehicle models without infrastructure changes.

Oil‑gas field stations : The robot can conduct long‑duration autonomous patrols, recognise anomalies and respond on‑site. Voice commands are translated into structured instructions; lidar, camera and force‑sensor data are fused to form a perception‑planning‑execution loop.

Port terminals : Multiple robots are coordinated to install or remove twist‑locks on containers. The brain schedules several robots in parallel, illustrating the “one brain, multiple bodies” concept.

Brain Architecture

The “RuoYu Jiutian” brain integrates perception, planning and execution in a single end‑to‑end model using large‑language models and 3‑D decoders. It follows a “brain‑cerebellum” hierarchy: the brain layer performs high‑level task planning via diffusion‑based imitation learning and 3‑D affordance perception; the cerebellum layer refines the output into joint‑level precise motions.

World‑Model‑Driven Prediction

In the fueling scenario the brain first predicts the future state of the environment (“target observation”) and then synthesises intermediate visual frames that satisfy both the coarse action plan and the desired end state. This enables anticipatory adjustments, such as repositioning before grasping the gun when the current angle would reduce success probability.

Target‑Driven Hierarchical Refinement (H‑GAR)

The H‑GAR framework proceeds in three steps:

Coarse action draft : Generate a rough sequence of actions from historical frames and the task command.

Goal‑conditioned observation synthesis (GOS) : Produce intermediate visual frames constrained by the target observation and the coarse draft.

Interactive perception‑action refinement (IAAR) : Upgrade the coarse actions to executable commands using feedback from the synthesized frames and a memory bank of past refined actions.

On the Libero‑10 multitask benchmark H‑GAR achieved a 94 % success rate. Real‑world long‑chain tasks (object placement, drawer opening) showed significantly higher stage‑completion rates than competing methods.

Multimodal Closed‑Loop Correction

The system continuously fuses visual changes with force feedback. When visual cues indicate a correct insertion but force data reveal abnormal resistance, the robot decides whether to fine‑tune the angle, retry, or backtrack, improving robustness in noisy outdoor conditions.

Generalizable “One Brain, Many Bodies” Design

A lightweight universal architecture allows the same brain to drive various robot morphologies (dual‑arm heavy‑load, single‑arm light‑load, etc.) and interchangeable explosion‑proof dexterous hands. Knowledge acquired in the fueling scenario—perception, planning, error correction—can be transferred to other domains without retraining from scratch, reducing integration cost and deployment time.

Paper

https://arxiv.org/pdf/2511.17079