Industry First: AiMOGA Robot Successfully Completes "Autonomous Opening of Car Door" Operation

Hefei, July 29, 2025 / Xinhua News Agency / -- Recently, AiMOGA's robot, Mornine, successfully completed the operation of "autonomously opening a car door" in a real-world 4S store environment, becoming the first robot to achieve this complex physical interaction task. This breakthrough not only marks the transition of embodied intelligence from the laboratory to the real world but also lays the foundation for the commercial deployment of service-oriented humanoid robots.

In the robotics industry, the seemingly simple "opening a car door" action actually requires extremely high demands on robot perception, planning, and control capabilities. The AiMOGA team through continuous engineering trials and algorithm iterations overcame multiple technical difficulties such as sensor fusion, whole-body control, and upper-lower limb coordination, ultimately enabling Mornine to possess the ability to autonomously recognize car doors, adjust its stance, coordinate movements, and open the door in non-structured scenarios.

Mornine Autonomous Opens the Car Door

Breaking the boundary of preset instructions, achieving end-to-end closed-loop control

Traditional robots often require step-by-step instructions or remote control to execute tasks, whereas Mornine adopts an end-to-end closed-loop control architecture, seamlessly integrating perception, planning, and execution. Based on multi-modal sensors such as 3D lidar, depth cameras, and wide-angle cameras, the robot builds a comprehensive environmental and self-state awareness model using visual foundation models (Visual Foundation Model) and leverages large-scale models for semantic understanding and task generation.

When faced with different car door handles or sudden resistances, Mornine can adjust its path, posture, and force output in real-time. This seemingly instinctive execution mechanism enables the robot to truly possess the ability to make immediate decisions and adjust actions in response to dynamic environments.

Strengthening Learning Enables Autonomous Recognition of "Door Handle" Targets

This autonomous opening capability is not achieved through human demonstration, but rather through hundreds of millions of virtual training iterations. During training, the robot only receives a reward signal for successfully grasping the door handle, yet it gradually focuses on the door handle region through repeated trial-and-error processes.

"We never explicitly told the robot what 'door handle' is, but it ultimately learned to grasp the target by itself." The research team noted. This ability is a direct manifestation of end-to-end reinforcement learning in real-world applications.

Attention Heatmap

Reinforcement Learning Training Model Self-Organizes to Focus on the Door Handle Region

Coupling Legs and Arms, Creating a "Human-Like" Action Chain

To address challenges in door handle grasping and stability, Mornine adopts a "flow-based motion (流动式动作机制)" mechanism. During the opening process, the robot not only uses arm force but also adjusts its leg and waist posture to form a whole-body coordination, simulating the structure of Tai Chi push hands.

The initial team attempted to rely solely on upper-limb force to open the door, but the robot was often pulled away by the door. Ultimately, through strengthening hand structures, increasing foot grip, setting reasonable posture and stance precision requirements, and optimizing control models, data feedback mechanisms, and strategy updating abilities, the team gradually achieved stable opening movements.

"From failure countless times to the first time opening the door, that moment we knew it had become a reality." The research team recalled.

Sim2Real Migration, Building from Virtual to Real-world Loop

The trained model migrates smoothly through Sim2Real technology to the physical robot, enabling virtual world strategies to be applied in real-world applications. During deployment, the robot can collect environmental feedback data and flow it back into the training system for model enhancement, forming a closed-loop.

This "Pre-training - Real-world Reinforcement - End-to-End Control" path is widely recognized as the most generic and scalable technology paradigm in the current robotics industry.

Robot Mornine Uses VLM Model and Wide-Angle Camera to Real-Time Judge Car Door Open Status

Landing in 4S Stores, Serving Real-world Users

Currently, Mornine has been deployed in multiple 4S stores, taking on intelligent reception, product explanations, and material delivery tasks. The "opening the door" is not only a technological breakthrough but also a key ability for its role as an "intelligent sales assistant."

In the future, AiMOGA will continue to accumulate data in automotive scenarios, optimizing model generalization capabilities. Faced with various door handles, lighting conditions, and opening resistances, the robot will evolve to generate new action strategies within minutes, significantly reducing the development cycle of new tasks.

"We opened one car door today, but we hope that robots can open more scenarios' 'service doors' in the future." From conceptualization to real-world applications, from laboratories to commercial spaces, AiMOGA's Mornine robot has completed the crucial leap in the commercial deployment of embodied intelligence.