Viam debuts BoxBot box-opening robot for IROS 2026

Viam announced BoxBot, a research arm that cuts tape and opens cardboard box flaps, pairing camera-guided cutting with a fine-tuned open-source vision-language-action model trained on 125 person-driven demonstrations collected in about five hours.

SOFTWARE desk — open demos plus an open vision-language-action model on a production robotics platform shrinks the gap between a person driving the arm in the lab and a policy that runs the same job on the robot.

Viam will debut BoxBot at IROS 2026 in Pittsburgh. The record is the company’s release on PR Newswire. The page stamp is Sep 22, 2026, 12:46 ET. The dateline is Pittsburgh. IROS is the robotics research conference the release names. The page says the meeting runs from September 27 through October 3, and it calls IROS one of the world’s leading robotics research conferences. Leading is the company’s word. This desk did not rank conferences. BoxBot is an ongoing Viam research project: a robot arm that opens taped cardboard boxes. These lines are Viam’s. This desk did not stand at the booth.

What the arm does, in the release’s split. The page says BoxBot specifies what it can and learns what it cannot. Image-based visual servoing cuts the tape. Visual servoing means the camera picture steers the tool while the cut is happening, instead of playing a path that was written in advance. The page says no two sealed boxes are alike, so the robot uses that image feedback to guide the cutting tip to tape seams that are typically 1 to 2 millimeters wide. A seam, here, is the taped line. A fine-tuned SmolVLA model then peels open and adjusts four box flaps. Fine-tuned means they trained an existing open model further on their own recordings. A vision-language-action model, shortened to VLA, looks at the camera and outputs the next movement. SmolVLA is the open VLA named on the page. The page calls the flap job deformable and contact-rich: the cardboard bends, and the hand has to stay against it. File both jobs as the release’s. This desk did not watch a box open.

Where those two jobs run. Both systems run as native services on the same machine. The VLA policy infers locally on an NVIDIA Jetson Orin Nano through Viam’s ML model service. Infer means the model is choosing the next move, not being trained. Locally means that choice happens on the small computer attached to the robot, not in a distant data center. A Jetson Orin Nano is the NVIDIA computer the page names for that step. A policy, here, is the trained behavior that drives the arm. These lines are the release’s. This desk did not time a move.

How the recordings were made. Two Viam engineers collected 125 demonstrations in about five hours, spread over four days, with a VR controller. A demonstration is one recorded run in which a person drove the arm through the task. VR, here, is a virtual-reality hand controller. The page says Viam synced the arm’s data and two camera streams to the cloud, the team fine-tuned SmolVLA with LeRobot on Hugging Face Jobs, and then put the policy back on the robot. LeRobot is the training tool named on the page. Hugging Face Jobs is the hosted place they ran that training. The page says there was no custom model architecture and no task-specific engineering: the base model and 125 episodes were enough for a working policy. An episode is one of those recorded runs. Enough, and working, are the company’s words. This desk did not rerun the training and did not count the hours.

The open files, and what they are not. The release says the code and the dataset are open at huggingface.co/viamrobotics, so researchers can train on them, benchmark against BoxBot, or reproduce the setup. A benchmark, here, would be a shared test another lab can score. The release does not print a success rate, a time per box, or a score against another robot. Do not add one. The Hugging Face organization page, read the same day, lists SmolVLA models named smolvla-box-bot and smolvla-box-bot-subtasks, and datasets named box-opener and box-opener-subtasks. Those names are the page’s. The release does not say which of those files is the booth policy. This filing does not pick one.

What the booth will show. The page says people at IROS can watch BoxBot run, drive the arm themselves, and record data on a separate station. Viam’s booth is 828. Three stations are named. BoxBot: the arm cuts the tape seam and opens the flaps. A VR teleoperation station uses the same VR setup that produced the training data. Teleoperation means a person is driving the arm. An SO-101 leader-follower station lets a person move a leader arm and watch a follower arm copy it, then record and replay an episode and see the synced data in the Viam app. Leader is the arm the person moves. Follower is the arm that copies. SO-101 is the station name on the page. The page does not define the arm’s brand. Do not add one. File the booth list as the release’s. This desk did not book a meeting.

The named voice, as color only. Shannon Bradshaw, Ph.D., senior vice president of product and education, said the question is not whether to use a VLA, it is where. She said learned policies are the right tool for contact-rich tasks that are hard to write as a formula, and that methods like visual servoing give precise answers when the scene keeps changing. She also said this robot uses classical motion planning for other parts of the job. Motion planning, here, means computing a path for the arm instead of learning that path from examples. She said many frameworks would need a pile of one-off scripts to combine those pieces, and that Viam puts the hardware, the software, and the models in one place the team can change. She said the implementation is public, so robotics and physical-AI researchers can check the work or build on it. Physical AI, here, means software that has to move a machine in the real world, not only answer on a screen. File the name, the title, and those sentences as the release’s. A quote is not a success rate.

Who Viam says it is. The about box calls Viam the software platform for building, deploying, and managing robotics applications. It was founded in 2020 by Eliot Horowitz, a co-founder and former chief technology officer of MongoDB. The headquarters and the robotics lab are in New York City. The page says the company works with teams in manufacturing, marine, climate, logistics, entertainment, and more. The software platform for robotics is the company’s line. This desk did not rank platforms. The page does not print a price, a customer for BoxBot, or a date when the arm would ship as a product you can buy. BoxBot is a research project on the way to a booth.

Leave the blanks blank. Do not turn a working policy into a rate this desk measured. Do not turn one of the world’s leading into a ranking this desk made. Do not turn the Hugging Face file names into a claim that one of them is the IROS policy. Do not name an arm brand the release left unnamed. Do not turn the booth into a warehouse that is already running. The contact lines for a meeting, and the note that developers connect hardware and write logic in more than one programming language, stay in Sources. They are not a second result.

Plain English for the rest of the card: BoxBot = Viam’s research arm that cuts tape and opens box flaps. IROS = the robotics research conference in Pittsburgh, September 27 through October 3. visual servoing = the camera steers the cutting tip while the cut happens. seam = the taped line, typically 1 to 2 millimeters wide on this page. VLA = a vision-language-action model, which looks at the camera and outputs a movement. SmolVLA = the open VLA they fine-tuned. fine-tuned = trained further on their own recordings. demonstration, or episode = one recorded run of a person driving the arm. 125 = how many of those runs the page says two engineers collected. about five hours over four days = the page’s time for that collection. VR = the hand controller they used. Jetson Orin Nano = the small NVIDIA computer on the robot that runs the policy. infer = the model choosing the next move. policy = the trained behavior. LeRobot = the training tool named on the page. Hugging Face Jobs = where the page says that training ran. leader = the arm a person moves. follower = the arm that copies it. booth 828 = where Viam says to find the demo. physical AI = software that has to move a real machine. This filing is the 22 Sep debut announcement. It is not a measured success rate.

PRIMARY here: Viam’s 22 Sep 2026 PR Newswire release, “Viam Debuts Box-Opening Robot at IROS 2026,” stamped 12:46 ET and datelined Pittsburgh — Tier A PRIMARY, the company’s own record. The Hugging Face organization page at huggingface.co/viamrobotics is the code host the release names, not a second newsroom. The IROS dates, the visual-servoing seam of 1 to 2 millimeters, the fine-tuned SmolVLA flap policy, the Jetson Orin Nano, the 125 demonstrations in about five hours over four days, the two engineers, the VR controller, the two camera streams, the LeRobot and Hugging Face Jobs line, the no-custom-architecture claim, booth 828, the VR station, the SO-101 leader-follower station, the Bradshaw title and quote, the 2020 founding by Eliot Horowitz, and the New York City headquarters and lab are the release’s. The model and dataset names on the organization page are that page’s. NOT claimed: a success rate, a time per box, an arm brand, that one named Hugging Face file is the booth policy, that this desk opened a box or reran the training, a product for sale, a stock tip, or investment advice. Distinct from the already-filed intrinsic-core-open-source, nvidia-isaac-ros-5, and boston-dynamics-rmac.