Flexible robotic manipulators that mimic the shape of an elephant's trunk can reach into tight spaces and grasp fragile objects. But they are hard to control: smooth bends make the next movement poorly predictable. Researchers from Peking University have proposed a method that keeps the computer model from "losing the thread" when predicting, even if the arm moves faster than it did in the training data.
The work was carried out by a group led by professors Xie Guangming and Zhang Feitian from Peking University's School of Advanced Manufacturing and Robotics. The results were presented at the Robotics: Science and Systems conference (RSS 2026), held in Sydney from July 13 to 17, 2026, and published in the proceedings Robotics: Science and Systems XXII.
Usually, before a flexible robot moves, the controller "rehearses" possible commands in a computer model and chooses the best one. Models trained on data predict one small step at a time, and each next step builds on the previous one. As a result, small errors accumulate: after 200 steps the prediction can diverge greatly from reality, and the robot will act on a mistaken idea of its own body.
The team's idea is to give the prediction a point of support. One model, called the "anchor", learns what shape the arm takes when the control signals remain unchanged. At each step it pulls the prediction toward that shape, preventing errors from growing. But the arm swings and bends on its way to rest, so a second, lighter model, the "correction", learns these dynamics from motion data. Together they form a hybrid model.
The method was first tested in simulation on a three-dimensional cable-driven arm. When it had to predict 6 seconds of motion, or 200 steps, which goes well beyond the training data, the hybrid model reduced the error by more than a quarter compared with the "anchor" alone. It also outperformed the standalone "correction" and never lost the trajectory in 19 trials.
Tests were then run on a real device: a hydraulic soft fish-like tail that swings from side to side under water pressure. Here the hybrid model's error was half or less compared with the "anchor" alone and a quarter or less compared with the "correction" alone. On standard test tasks it completed all predictions, whereas the competing method sometimes failed.
More robust predictions give the robot's controller something to rely on. It can first try out many possible movements on the computer and choose the safest before the real arm starts moving. A robot able to foresee how its own body will bend can plan a careful movement in advance. This could help flexible manipulators work where rigid machines struggle to reach, for example in tight spaces and with easily breakable objects.
The researchers stress that this concerns a statistical association between model architecture and prediction accuracy, not a proven causal effect. The hybrid approach was tested in simulation and on one real device; its applicability to other types of flexible robots remains to be studied.