ROBOTSIDER
News · 2026-09-25 · Robotsider

Viam BoxBot Opens Boxes With a Hybrid Stack

Viam will show BoxBot at IROS 2026, pairing visual servoing for tape cutting with a SmolVLA policy to open box flaps.

Viam BoxBot is a research robot that Viam says will open taped cardboard boxes at IROS 2026 in Pittsburgh. The system uses camera-guided visual servoing to cut tape, then a fine tuned SmolVLA vision-language-action policy to peel open the flaps.[1]

Viam says two engineers recorded 125 teleoperated demonstrations in about five hours across four days. The company describes BoxBot as an ongoing research project, not a commercial product, and did not publish a task-success rate, cycle time, price, or deployment plan.[1]

Diagram of Viam BoxBot's hybrid control stack, with visual servoing for tape cutting and a fine tuned SmolVLA policy for flap opening, feeding local Jetson Orin Nano inference.
BoxBot divides the box-opening task between visual servoing and a learned policy. Source: Viam, September 22, 2026.[1]

How does Viam BoxBot open a taped box?

Viam BoxBot assigns the precise cut to visual servoing and the deformable flap work to a learned policy. Viam says image feedback guides the cutting tip to tape seams 1 to 2 mm wide. A fine tuned SmolVLA model peels and adjusts four flaps.[1]

That division matters because the two parts of the task fail differently. Tape cutting needs position feedback near a narrow seam. Pulling apart cardboard flaps involves contact and changing geometry. Viam says it paired those methods instead of using one learned controller for the full sequence.[1]

"Learned policies are the right tool for contact-rich manipulation tasks that are hard to model analytically," said Shannon Bradshaw, Viam's SVP of Product and Education.[1]

SmolVLA is an open-source, 450 million parameter vision-language-action model from Hugging Face. Its developers describe task-specific post-training from robotics data.[2] That supports Viam's choice of a task-specific fine tune. It does not establish BoxBot's performance beyond the company's demonstration.

For background on middleware choices around robot software, see our ROS 2 Lyrical versus Jazzy guide.

What did Viam disclose about the training and deployment?

Viam says its engineers captured synchronized arm data and two camera streams with a VR controller. The team then fine tuned SmolVLA with LeRobot and deployed the policy to a Jetson Orin Nano through Viam's ML model service.[1]

Viam's platform documentation says its ML service runs inference on the machine's hardware. It supports local-first operation without a network connection.[3] The announcement says BoxBot used local Jetson inference. It does not provide latency, compute use, recovery behavior, or a comparison with an all-learned alternative.[1]

The company says its implementation is publicly available. Its Hugging Face organization page showed no public models or datasets when checked for this article.[1] That gap may be temporary. The cited materials do not currently let a reader reproduce the 125-demonstration claim.

Warehouse teams considering where manipulation automation fits can compare this research setup with current humanoid warehouse deployments.

When and where will BoxBot appear?

Viam says it will show BoxBot at IROS 2026, scheduled for September 27 through October 3 in Pittsburgh. The company says attendees can watch the system cut tape and open flaps, use a separate VR teleoperation station, and see a leader-follower arm record and replay an episode.[1]

The demonstration can show whether the stated hybrid stack works on the boxes Viam brings to the booth. It cannot answer every operational question a buyer needs for a box-opening cell. Those include acceptable box variation, intervention frequency, safety controls, throughput, and service requirements. Viam has not announced those measurements.[1]

Viam BoxBot FAQ

Frequently asked questions

What is Viam BoxBot?

BoxBot is Viam's research robot for opening taped cardboard boxes. Viam says it uses visual servoing to cut tape and a fine tuned SmolVLA policy to open the flaps.

How much training data did Viam use for BoxBot?

Viam says two engineers collected 125 teleoperated demonstrations in about five hours over four days.

Is Viam BoxBot a commercial box-opening product?

Viam calls BoxBot an ongoing research project. Its announcement does not give a price, task-success rate, cycle time, or commercial availability date.

Where will Viam demonstrate BoxBot?

Viam says it will demonstrate BoxBot at IROS 2026 in Pittsburgh from September 27 through October 3.

Sources

  1. https://www.prnewswire.com/news-releases/viam-debuts-box-opening-robot-at-iros-2026-302886481.html
  2. https://huggingface.co/blog/smolvla
  3. https://www.viam.com/platform/ai-and-data