A Deep Reinforcement Learning approach to Robot Grasping

bg-shape illustrations illustrations illustrations illustrations illustrations
A Deep Reinforcement Learning approach to Robot Grasping

Date

Jul 30, 2020

Author's

Ashutosh Tiwari

Categories

Robotics, reinforcement learning

Paper details

Under Revision in Robotics adn Autonomous Systems (RAS), Elsevier Journals

  • Trained a Robotic Arm to pick up and get a stable grip on an non-deformabale arbitrary object.
  • Developed an RL model exploring PPO, DDPG by varying different feature learning methods and custom input structures
  • Evaluated speed performance in training time (~50,000 episodes) and achieved overall accuracy (~70%) with PPO trained on a combination of random features from a fixed CNN and feature vectors