A New Method Combining Single-Point Push and Double-Point Complete Push for Partially Observable Scenarios

Pushing objects into a target configuration is an important skill for robots. When there are obstacles in the scenario, the movement range of the objects will be limited, and objects are easily lost due to obscured by obstacles, which complicates the pushing task. This paper proposes a push path planning algorithm SDP-RRT (Single and Double Push - Rapidly-exploring Random Tree) to combine single-point push and double-point complete push in a partially observable environment with obstacles. The single-double push combines the flexibility of a single-point push and the stability of a double-point complete push. To find the reliable pushing point and pushing direction of the double-point complete push, we improve the judgment conditions of the double-point complete push and estimate the object's center of mass based on the LSTM (Long Short-Term Memory) neural network. Finally, the success rate and push steps of single-double and single-point push are compared in both the simulation and the natural environments, and the feasibility of the SDP-RRT algorithm is verified.