Reinforcement Learning for Robotic Object Picking

Deep reinforcement learning for vision-based object picking with a KUKA manipulator using DDPG and A3C.

DDPG training for object picking with a KUKA manipulator
A KUKA manipulator learns to grasp objects through trial-and-error using deep reinforcement learning.

This project investigates deep reinforcement learning for robotic manipulation, where a simulated KUKA manipulator learns to grasp objects through interaction with its environment. The objective is to train policies capable of reliably reaching, grasping, and lifting objects without manually designing the control strategy.

Deep Deterministic Policy Gradient (DDPG)

The initial implementation uses Deep Deterministic Policy Gradient (DDPG), an off-policy actor-critic algorithm designed for continuous control. The actor network learns a deterministic control policy while the critic estimates the expected return of state-action pairs. Training is stabilized using replay buffers, target networks, and exploration noise, allowing the robot to gradually improve its manipulation strategy through experience.

Highlights

  • Implemented a complete DDPG training pipeline in PyTorch
  • Trained a KUKA manipulator to perform object picking in simulation
  • Implemented actor and critic neural networks with target network updates
  • Used experience replay and exploration noise for stable continuous-control learning
  • Evaluated training through reward, episode length, and loss metrics