Humanoid robot combines grasping and walking with 93% zero-shot success

Researchers Fanxing Meng and Jing Xiao present a method for humanoid robots to learn whole-body reaching and grasping policies from scratch, then combine them with separately learned locomotion policies to perform complex sequences like grasping, standing up, and walking. The grasping policy uses cubic harmonics—a mathematical representation from harmonic analysis—to model the spatial relationship between the hand and object via spatial convolution. This enables the robot to generalize to unseen objects without external datasets or pretrained models. A developmental curriculum that decouples finger joints during training further improves learning efficiency and generalization.

The combined policy integrates the grasping policy with a standing-up and walking policy by providing each with its own observation vector and using hand-object interaction scores to decide which policy should control which robot joints at any given moment. The results show a 93% zero-shot success rate for grasping unseen objects and a 96-100% success rate for standing up while holding an object. Notably, the study demonstrates that combining policies only works if both are learned on the same whole humanoid body—even when one policy, like locomotion, might not seem to depend on all body parts such as fingers. This finding has significant implications for robot skill acquisition and developmental robotics.