Viam’s BoxBot combines visual servoing and SmolVLA for robotic box opening

Date: 2026-09-22
Type: ARTICLE
Desk: Robotics
Topics: Packaging Automation, Robot Learning, Vision Language Action Models, Visual Servoing
Deck: The research system uses image feedback for tape cutting and a locally deployed policy for box flaps, trained from 125 teleoperated demonstrations.
URL: https://korthosrobotics.com/research/viam-boxbot-visual-servoing-smolvla-box-opening-2026
Referenced profiles: Viam (https://korthosrobotics.com/ecosystem/company/viam)

Image: Viam
Viam introduced BoxBot on 22 September 2026, describing a research robot that cuts the tape on sealed cardboard boxes and opens their flaps. The system separates the task into image-based visual servoing for precise cutting and a fine-tuned SmolVLA policy for deformable, contact-rich flap manipulation, with a demonstration planned at IROS.
Image feedback guides the cutting tip toward tape seams typically 1–2 mm wide, while the learned policy peels open and adjusts four flaps. Both run as native services on the robot, with VLA inference performed locally on an NVIDIA Jetson Orin Nano. Classical motion planning handles other elements of the application alongside the visual and learned controllers.
Two engineers used a VR controller to collect 125 demonstrations in about five hours over four days. Viam synchronised arm data with two camera streams, after which the team fine-tuned SmolVLA using LeRobot on Hugging Face Jobs and deployed the policy back to the machine. The Viam Robotics repository collection provides the released code and dataset for reproducing the setup or testing alternative policies.
Viam’s earlier November 2025 partnership with Universal Robots applied the same software platform to industrial processes including fibreglass sanding, gelcoat and paint application, wood sanding and polishing. Those systems use UR-series arms, while Viam’s New York and Italian laboratories test hardware from multiple manufacturers.