Improved Exploration through Latent Trajectory Optimization in Deep Deterministic Policy Gradient

Model-free reinforcement learning algorithms such as Deep Deterministic Policy Gradient (DDPG) often require additional exploration strategies, especially if the actor is of deterministic nature. This work evaluates the use of model-based trajectory optimization methods used for exploration in Deep Deterministic Policy Gradient when trained on a latent image embedding. In addition, an extension of DDPG is derived using a value function as critic, making use of a learned deep dynamics model to compute the policy gradient. This approach leads to a symbiotic relationship between the deep reinforcement learning algorithm and the latent trajectory optimizer. The trajectory optimizer benefits from the critic learned by the RL algorithm and the latter from the enhanced exploration generated by the planner. The developed methods are evaluated on two continuous control tasks, one in simulation and one in the real world. In particular, a Baxter robot is trained to perform an insertion task, while only receiving sparse rewards and images as observations from the environment.

Authors

• Kevin Sebastian Luck
• Mel Vecerík
• Simon Stepputtis
• Heni Ben Amor
• Jonathan Scholz

Keywords

• Training
• Symbiosis
• Heuristic algorithms
• Computational modeling
• Performance gain
• Deep reinforcement learning
• Probabilistic logic
• Linear programming
• Safety
• Trajectory optimization
• Latent Trajectory
• Deep Learning
• Value Function
• Reinforcement Learning Algorithm
• Exploration Strategy
• Deep Reinforcement Learning Algorithm
• Model-free Reinforcement Learning
• Image Embedding
• Objective Function
• Dynamic Network
• Autoencoder
• Latent Space
• Functional Dynamics
• Actor Network
• Subsequent Imaging
• Reward Function
• Latent State
• Gradient-based Optimization
• Ornstein-Uhlenbeck Process
• Planning Horizon
• Critic Network
• Gradient-based Optimization Methods
• Reinforcement Learning Methods
• Encoder Network
• Cheetah
• Deterministic Methods
• Policy Search
• Contrastive Loss