Robot Grasp Learning on the Moon
Robot Grasp Learning on the Moon from 3D Observations
Robot Grasp Learning on the Moon
Robot Grasp Learning on the Moon from 3D Observations
Space rovers equipped with a general-purpose robotic arm have numerous applications in lunar and planetary exploration. The addition of autonomy to such systems is desirable in order to increase the amount of time rovers can spend gathering scientific data and collecting samples.
This project looks into the use of Deep Reinforcement Learning for vision-based robotic grasping of objects on the Moon. A novel simulation environment with procedurally generated datasets is developed to train agents in unstructured scenes with uneven terrain and harsh lighting.
The policy is then end-to-end learned using a model-free off-policy actor-critic algorithm that directly maps compact octree observations to continuous actions in Cartesian space. When compared to traditional image-based observations, experimental results show that 3D data representations enable more effective learning of manipulation skills.
Domain randomization improves the generalization of learned policies to novel scenes with previously unseen objects and different lighting conditions. To that end, we demonstrate zero-shot sim-to-real transfer by evaluating trained agents on a real robot in a Moon-analogue facility.
Project Facts
PROJECT NAME
Robot Grasp Learning on the Moon from 3D Observations
EFFECTIVE START/END DATE
September 2021 - .. ongoing
PROJECT PARTNERS
- Aalborg University
- Luxembourg University
AAU Space Group