Robots can rehearse handling soft things, such as plush toys, in a simulated world, but only if the virtual objects squash and spring back the way real ones do. Nankai University and Rightly Robotics have built DeformSmith, which makes such objects automatically from a short text description or a single image. How an object gives under pressure affects whether a robot can grab it, carry it and let it go, and a picture alone shows little of that.
DeformSmith works like a small workshop with a strict inspector. A team of AI helpers shapes the object and decides how heavy and how soft it should be, then a built-in physics tester drops it, squeezes it and watches how it recovers. Next, a simulated robot arm tries to pick it up, carry it and set it down, and if something goes wrong, the system adjusts the robot's moves or the object's material. Each change is checked again before the object is accepted.
What comes out is a ready-to-use virtual object, plus replayable recordings of the robot handling it, which robot teams could use as practice material. In a blind side-by-side test, 40 people mostly preferred DeformSmith's objects over those from rival methods, and, leaving out ties, picked them about nine times in ten for looks. When dropped, its virtual turtle, whale and seal kept their shape, while some rival versions flattened or fell apart. The team also found that feeding the robot's practice runs back into the design led to more successful pickups.
The team says its code will be released, along with more demo videos, but its pages give no date. For now, DeformSmith handles solid objects made of one evenly soft material; liquids and loose, grainy materials are outside its scope, and the robot's contact with objects is only roughly modeled. Next, the researchers want to use video of real objects to adjust how soft each virtual object is, and to support a wider range of materials.
It is not yet clear how closely these virtual objects match real ones. The tests judged how convincing they looked and moved, not whether their softness matches a real toy, and the team says settings guessed from text or a picture still need checking in the real world.