ROMA Teaches Robots to Touch, Shake and Squeeze Objects to Understand Them

Handed an unfamiliar box, a person rarely stops at looking. They pick it up to gauge the weight, shake it to hear whether something rattles inside, and squeeze it lightly to tell whether it is soft or rigid. Three simple actions yield three different physical parameters. ROMA, which the Chinese outlet Jiqizhixin wrote about on October 10, 2026, is built on this observation: a robot should not passively accept camera signals. It should notice on its own that it knows too little about an object and gather the missing information through interaction.

We have only the introductory part of the piece, so what follows covers what it says and what remains unknown.

What ROMA claims

The core thesis is stated clearly: human perception of the world is not limited to receiving ready-made sensory data. When we realize information is missing, we act, and the feedback we get completes the picture. ROMA carries this logic over to a robot. The system is supposed to explore an object through physical contact rather than rely only on what is visible in an image.

What matters here is not the hand movement itself but the condition that triggers it. The robot first assesses what it is missing, and only then chooses the action that will close the gap. This framing rules out a "touch everything" scenario from the start.

Why vision alone is not enough

Mass, stiffness, friction and whether an object has contents are hard to see in a frame. Two boxes that look identical can weigh 100 grams and 2 kilograms. Modern vision-language models describe a scene well, but they can only estimate an object's physical properties from a picture approximately. For a robot that has to grasp an item without crushing it, dropping it or spilling its contents, that imprecision is often not enough.

Interactive perception is not a new idea in robotics: researchers have worked on it for decades, and tactile sensors such as GelSight have long been used to estimate shape and texture. The question is who controls the exploration. In earlier work, the sequence of actions was often set in advance. ROMA, judging by the description, makes the choice of action part of the agent's own reasoning.

What the available excerpt tells us

• Source: Jiqizhixin (机器之心), published October 10, 2026.
• Name of the approach: ROMA; the task is interactive object exploration by a robot.
• Human analogues cited in the piece: weighing, shaking, squeezing.
• Logic: interaction is triggered when the system detects a lack of information.
• Metrics, the list of test objects, and the type of manipulator and sensors are not disclosed in the available text.

What is missing for an assessment

Without experimental data, it is too early to judge how strong the approach really is. It is unknown what hardware the system was tested on, which objects it handled, or how much the grasp success rate or the accuracy of mass estimation improved over baseline methods. It is also unclear how ROMA relates to large multimodal models: whether it is trained from scratch, uses a language model as a "planner," or is fine-tuned on interaction trajectories.

What developers should watch for

When the full paper or a repository appears, four things are worth checking. First, how the "lack of information" is formalized: through model uncertainty, a separate classifier, or language-based reasoning. Second, how many physical attempts are spent on a single object, because every extra shake costs cycle time and mechanical wear. Third, whether the policy transfers to other manipulators and sensors. Fourth, whether there is a comparison against a simple "always lift and weigh first" scheme: if that works almost as well, the value of choosing actions will have to be proven separately.

For teams building warehouse and household robots, ROMA is for now a signpost for direction, not a ready-made tool. The direction is clear, though: physical understanding of objects is gradually moving from the camera to the hand.