Legged robots have the potential to leverage obstacles such as rocks and boulders to climb steep sand slopes. However, efficiently repositioning these obstacles to desired locations remains challenging. Here we present DiffusiveGRAIN, a learning-based method that enables a multi-legged robot to strategically induce mini sand avalanches using their legs, indirectly manipulating obstacle positions. Using a laboratory granular trackway, we perform 300 loco-manipulation experiments with systematically varied obstacle distance, robot orientation, and robot actions. Experimental data reveal that when obstacles are relatively separated (i.e., distance > one obstacle diameter), their movements under leg-induced granular flow are largely independent. However, for closely-located obstacles, granular flow interference between them becomes significant, and their movement needs to be resolved jointly. In addition, we find that different robot actions can yield distinct changes in robot state and impact subsequent terrain manipulation options, requiring joint prediction and planning of manipulation and locomotion actions. To address these challenges, DiffusiveGRAIN includes a diffusion-based “environment predictor” to capture multi-obstacle movements under granular flow interferences. In addition, we develop a “robot state predictor” to estimate changes in robot state from various leg action patterns. Experiments demonstrate that by integrating the environment and robot state predictors, a multi-legged robot can autonomously plan its leg movements based on loco-manipulation goals, successfully shifting closely located rocks and boulders to desired locations in over 70% of trials. Our study showcases the potential for multi-robot teams to collaboratively manipulate obstacles to achieve improved mobility on challenging terrains. 

For leg manipulation, we collect 240 (60 × 3 + 60) trials, which result in a total of 24,590 images. Specifically, we conduct 60 trials for each gantry manipulator manipulating obstacles at orientations of 0, 15, and 30 degrees, with both the left and right manipulators executing excavation actions. We also do 10 trials each for the gantry manipulator manipulating obstacles at orientations of 0, 15, and 30 degrees with just the left or right leg. The time interval between two consecutive excavation actions is 3 s. Before each trial, a human operator manually smoothes the granular media to a (roughly) even granular slope with an inclination angle 20 degrees. This inclination angle is close to the angle of repose of our granular material. Once the human prepares the granular surface, he/she places multiple 3D-printed (PLA) obstacles on the granular surface at different locations relative to the leg. For all 240 trials, we use a semi-spherical obstacle with diameter 5 cm.