How MTS Is Teaching a Humanoid with a VLA Model to Grab Objects Moving on a Conveyor

On July 22, 2026, MTS ML engineer Daniya published a write-up on the company's corporate blog on Habr describing how the company's Physical AI R&D unit is teaching a full-size humanoid to grab objects moving along a conveyor using a VLA policy (Vision-Language-Action).

What is the main problem?

The humanoid confidently picks up an object from a table, but starts missing when the same object rides a conveyor. That is the exact task the MTS team is focused on. In a static scene the object stays still, the lighting is stable, and the camera points at the work area, so modern VLA models deliver high-quality control. As soon as the object starts moving, the usual grasping accuracy disappears: a model trained on stationary scenes cannot keep up with the changing geometry of the task.

According to the engineer, this is not an isolated failure but a systemic limitation of an approach that works well in lab demos and transfers poorly to a real conveyor.

The robot confidently picks up an object from a table, but starts

missing when the object rides a conveyor.

— Daniya, ML engineer at MTC Web Services

Why is a dynamic scene harder than a static one?

A dynamic environment adds two layers of difficulty at once: the object itself moves, and so does the robot, which has to adapt to it. Most impressive robot demos are still filmed in static settings, with objects neatly laid out on a table and a scene that barely changes. Real tasks at home and in production do not look like that: at the very least, objects ride conveyors, and the grasp has to account for their speed and trajectory.

Key project details:

• Team: the Physical AI R&D unit at MTC Web Services
• Technology: a VLA policy (Vision-Language-Action) for a humanoid
• Author of the write-up: Daniya, ML engineer at MTS
• Focus: grasping objects moving along a conveyor, not objects lying on a table
• Publication date of the write-up: July 22, 2026

Why is it harder for a humanoid than for a manipulator?

A full-size robot has to do more than reach for an object: it must also keep its whole body balanced. A fixed manipulator solves a simpler problem. Its base is anchored, and the entire model is responsible only for arm movement. To make the same grasp, a humanoid has to take steps and compensate for shifts in its center of gravity with additional torso movements.

Every such movement changes the position of the camera and the hand relative to the object, which is itself moving at that moment. As a result, motion planning becomes many times more complex: the model has to predict the object's trajectory, plan the grasp, and keep the robot stable all at once. This is the combination the MTS team is now trying to assemble within its Physical AI unit.

What does this mean?

Moving from static demos to working with moving objects is one of the main barriers between humanoids in the lab and humanoids in real production. The MTS write-up shows that even the basic "pick up an object" operation stops being trivial once the scene comes alive, and that industrial use of VLA robots is limited not by grip strength but by the dynamics of the environment.

Frequently asked questions

What is a VLA policy?

VLA (Vision-Language-Action) is a robot control model that links vision, a language instruction, and action: the robot "sees" the scene, understands the task, and plans a movement. In static scenes, according to the MTS write-up, such models deliver high-quality grasping.

Why is a dynamic environment harder for a humanoid?

Because both the object and the robot itself are moving. To grasp, a full-size humanoid has to take steps and keep its body balanced with additional movements, so motion planning becomes many times more complex than for a fixed manipulator.