Viam plans to demonstrate BoxBot, a robotic arm designed to open taped cardboard boxes, at the IROS 2026 robotics conference in Pittsburgh. BoxBot is an ongoing research project that combines image-based visual servoing with a vision-language-action model. The visual servoing system uses camera feedback to guide a cutting tool along tape seams, which Viam said are typically 1 to 2 millimeters wide. A fine-tuned SmolVLA model handles the subsequent task of opening and adjusting the box’s four flaps.
The two systems operate as services on the same machine. The VLA model performs inference locally on an Nvidia Jetson Orin Nano using Viam’s machine-learning model service. The robot also uses conventional motion planning for other parts of the application.
Viam said two of its engineers collected 125 demonstrations for training the system over about five hours across four days. Using a virtual-reality controller, the engineers generated demonstrations while Viam’s software synchronized robotic-arm data with feeds from two cameras.
The company then fine-tuned SmolVLA using LeRobot on Hugging Face Jobs and deployed the resulting policy to the robot. Viam said the project did not require a custom model architecture. The company has made the BoxBot code and dataset publicly available on Hugging Face for researchers seeking to reproduce or benchmark the system.
At IROS, Viam plans to operate BoxBot alongside two additional demonstration stations. One will allow attendees to control a robotic arm using the VR setup employed to generate BoxBot’s training data. Another will use an SO-101 leader-follower configuration to demonstrate arm mirroring, episode recording and replay, and synchronized data capture.
Shannon Bradshaw, Viam’s senior vice president of product and education, said the project was intended to combine learned policies with methods such as visual servoing and classical motion planning within a single robotic application.
Viam’s platform provides hardware abstraction, computer vision, data capture, visual servoing and edge machine-learning deployment through a common set of APIs. Developers can configure hardware using JSON and write application logic in Python, Go or TypeScript. The platform can also operate alongside ROS 2-based robotic systems through Viam’s ROS 2 integration module.
