Learning Granular Media Avalanche Behavior for Indirectly Manipulating Obstacles on a Granular Slope
Haodi Hu, Feifei Qian*, Daniel Seita*
University of Southern California
* = equal advising
Learning Granular Media Avalanche Behavior for Indirectly Manipulating Obstacles on a Granular Slope
Haodi Hu, Feifei Qian*, Daniel Seita*
University of Southern California
* = equal advising
Overview
Abstract
Legged robot locomotion on sand slopes is challenging due to the complex dynamics of granular media and how the lack of solid surfaces can hinder locomotion. A promising strategy, inspired by ghost crabs and other organisms in nature, is to strategically interact with rocks, debris, and other obstacles to stabilize movement. To provide legged robots with this ability, we present a novel approach that leverages avalanche dynamics to indirectly manipulate objects on a granular slope. We use a Vision Transformer (ViT) to process image representations of granular dynamics and robot excavation actions. The ViT predicts object movement, which we use to determine which leg excavation action to execute. We collect training data from 100 real physical trials and, at test time, deploy our trained model in settings not seen in training. Experimental results suggest that our model can accurately predict object movements, and can also generalize to objects with different physics properties. To our knowledge, this is the first paper to leverage granular media avalanche dynamics to indirectly manipulate objects on granular slopes.
We propose Granular Robotic Avalanche INteraction (GRAIN), a novel learning-based method for leveraging granular avalanche dynamics for indirectly manipulating objects on a granular slope. Due to a lack of accurate simulators for simulating legged robots and avalanche behavior on granular slopes, we do all experiments in the real world. We use an RHEX family robot leg as an external disturbance source which performs excavation actions within a grain tank with mechanical support to form a slope. We also design a gantry structure to enable the robot leg to move to different positions for performing leg excavations. We collect training data through physical interaction and deploy our trained model in the real world with a quadrupedal robot.
Data Collection Video
Third-person view of the data collection process
An RGBD camera is mounted on top of the tank to collect images
Experiment Videos
Single obstacle with single task manipulation
GRAIN (Success)
Vector representation (Failure)
Baseline (Failure)
Red shaded areas are targets, green dots are the prediction of post-excavation locations of obstacles at the current step.
Single obstacle with sequential tasks manipulation
GRAIN (Success)
Vector representation (Failure)
Baseline (Failure)
Red shaded areas are targets, green dots are the prediction of post-excavation locations of obstacles at the current step.
Multiple obstacles manipulation
GRAIN (Success)
Vector representation (Failure)
Baseline (Failure)
Red shaded areas are targets, green dots are the prediction of post-excavation locations of obstacles at the current step.
Unseen obstacle manipulation
GRAIN (Success)
Vector representation (Failure)
Baseline (Failure)
Red shaded areas are targets, green dots are the prediction of post-excavation locations of obstacles at the current step.
Quadrupedal robot experiment
GRAIN (Success)
GRAIN (Success)
Vector representation (Failure)
Baseline (Failure)
Red shaded areas are targets, green dots are the prediction of post-excavation locations of obstacles at the current step.
Multiple unseen obstacles manipulation
GRAIN (Success)
Red shaded areas are targets, green dots are the prediction of post-excavation locations of obstacles at the current step.
BibTex
@inproceedings{hu2024grain,
title = {{Learning Granular Media Avalanche Behavior for Indirectly Manipulating Obstacles on a Granular Slope}},
author = {Haodi Hu and Feifei Qian and Daniel Seita},
booktitle = {Conference on Robot Learning (CoRL)},
Year = {2024}
}