video: In real-world experiments, the trained policy controls a dexterous robotic hand to twist objects with different geometries and physical properties, including screwdrivers, valves and light bulbs, without object-specific fine-tuning.
Credit: ©Science China Press
Twisting a bottle cap, turning a valve or screwing in a light bulb are simple actions for human hands. For robots, however, these everyday motions remain difficult. They require a hand to grasp, stabilize, slide and apply torque while adapting to objects with very different shapes, sizes, surface friction and stiffness.
A team from Huazhong University of Science and Technology and collaborators has developed DexAnyTwist, a learning framework that enables a dexterous robotic hand to acquire general twisting skills. The study reframes twisting as a hybrid manipulation problem. Instead of forcing one neural policy to average across all object types, the framework identifies different physical regimes and learns specialized expert policies for them.
To train and test the system, the researchers built a large library of twistable objects containing more than 300 instances across 10 categories, including bottles, nuts, rotation switches, shampoo bottles, liquor bottles, valves, cosmetics, screwdrivers and light bulbs. The dataset covers a broad range of everyday and industrial objects that require functional twisting, rather than simple grasping.
The study shows that the difficulty is not merely a lack of data. As object diversity increases, a single monolithic policy can degrade because different objects involve different contact dynamics. A smooth cap may depend heavily on friction, while a nut or screwdriver may require different contact patterns and geometric constraints. When one policy tries to fit these incompatible physical rules, learning can collapse into an averaged strategy that performs poorly on difficult cases.
DexAnyTwist addresses this challenge through Iterative Hybrid Manipulation Identification. The algorithm progressively partitions the task into dynamically consistent subsystems. At each stage, it identifies objects that one expert can handle reliably, removes them from the remaining difficult set, and trains a new expert on the residual objects. A gating mechanism then selects the most suitable expert according to the object geometry and current manipulation context.
The framework combines reinforcement learning in simulation with geometry-aware perception and expert routing. Object point clouds are encoded to capture shape information, while the policy uses hand and object states to respond to contact-rich interactions over time. This design allows different experts to specialize in distinct twisting modes without blending incompatible actions into a single average command.
In simulation, DexAnyTwist achieved the best overall performance among the tested methods. On the full object set, it reached a 90.4% success rate. The advantage was clearest on the hardest 30% of objects, where DexAnyTwist achieved a 69.5% success rate, compared with 51.0% for the strongest baseline. It also more than doubled the consecutive success count of the closest competitor in this hard subset, indicating stronger long-horizon stability.
The learned policy also transferred to real-world objects without object-specific fine-tuning. Deployed on a dexterous robotic hand, the system successfully twisted representative objects such as valves, screwdrivers, nuts and light bulbs. It consistently outperformed comparison methods in real-world tests, including on objects with narrow contact surfaces or fragile curved bodies. The real-world controller connected high-level policy outputs with fast low-level motor control, helping the learned skill operate on physical hardware. The experiments also tested objects with different materials and geometries, showing that the method can adapt through closed-loop feedback rather than object-specific labels.
Beyond task success, the framework developed interpretable strategies. Fingertip trajectory analysis showed that the thumb acted as the main driver of rotational torque, while the index and middle fingers stabilized the object as virtual pivots, resembling a human-like division of labor. In a light bulb experiment, the policy began with thumb-dominated motion and increased index-finger force as the bulb tightened, showing adaptive response to changing torque and friction.
The authors note that the current system still requires manual specification of the initial pose between the hand and object, and it is not yet optimized for high-torque operations such as opening tightly sealed mechanisms. Even so, DexAnyTwist points toward a scalable route for robots to learn contact-rich functional skills. Rather than only grasping objects, future robots may use such approaches to manipulate diverse objects in everyday environments.