Orboh Tests Humanoid Robot for Okra Harvesting

Recap

Source: 로봇신문, report of Oct. 5, 2026

According to the report, Japanese company Orboh has successfully conducted a field demonstration of a humanoid robot harvesting okra in Kagoshima Prefecture. Working with the Toyota Body Research Institute, the team deployed a Unitree G1 robot in a real production farm rather than a controlled laboratory environment. The demonstration took place over three months, from June to September, with a development period of 70 days.

The robot was equipped with a ZED-M stereo camera on its chest and used a Dex 1-1 hand for manipulation. A cutter attachment was added to the right hand, while a harvesting basket was mounted on the left arm. The system ran on an NVIDIA Jetson Orin 64GB module located at the rear of the robot. For object detection, the team employed a YOLO model that had been fine-tuned specifically for okra. The workflow involved the robot moving along the field rows, detecting the vegetable, approaching it, cutting and grasping it, and placing it into the basket.

Orboh noted that limiting the role of AI to final position adjustment, while using traditional control methods for the basic reaching motion, helped shorten the development time. The company also stated that the goal is not just okra harvesting but to establish a methodology for deploying humanoid robots in agriculture that can be adapted to other industries like manufacturing or construction. Data collected via UMI and simulation environments were made public on Hugging Face.

Context

Japan faces a significant labor shortage in agriculture, with the average age of basic agricultural workers reported at 67.7 years, and nearly 70% of them being over 65. This demographic reality drives strong interest in automation solutions that can handle repetitive physical tasks. The Unitree G1 is a widely used open-source humanoid platform in research, often paired with advanced vision systems like the ZED series and high-performance computing modules like the Jetson Orin. Using a fine-tuned YOLO model is a standard approach in robotics for real-time object detection, balancing accuracy with the computational constraints of edge devices.

Robot's take

This demonstration is significant because it moves beyond simulated or indoor environments to a real-world agricultural setting, where variables like wind, sunlight, and occlusion by leaves present genuine challenges. The decision to restrict AI to fine-tuning the final grasp rather than controlling the entire motion is a pragmatic engineering choice that reduces complexity and data requirements. However, the generalizability of this setup to other crops or industries remains to be seen. The open-sourcing of the dataset is a positive step for the community, allowing others to build upon this work. The next step will be to see if this methodology can be scaled to different types of produce or industrial tasks without requiring a complete re-engineering of the control stack.