Optimizing the Long-Term Behaviour of Deep Reinforcement Learning for Pushing and Grasping
This work addresses a specific limitation in robotic manipulation tasks like pushing and grasping, where existing methods struggle with long-term planning, making it an incremental improvement focused on domain-specific applications.
The paper tackled the problem of deep reinforcement learning systems failing to accurately predict long-term rewards in tasks requiring high foresight, such as a new bin sorting task, and found that combining the Hourglass model with Double Q-Learning enabled accurate long-term action sequence predictions, with results showing this technique is essential to prevent Q-value divergence when using high discount factors.
We investigate the "Visual Pushing for Grasping" (VPG) system by Zeng et al. and the "Hourglass" system by Ewerton et al., an evolution of the former. The focus of our work is the investigation of the capabilities of both systems to learn long-term rewards and policies. Zeng et al. original task only needs a limited amount of foresight. Ewerton et al. attain their best performance using an agent which only takes the most immediate action under consideration. We are interested in the ability of their models and training algorithms to accurately predict long-term Q-Values. To evaluate this ability, we design a new bin sorting task and reward function. Our task requires agents to accurately estimate future rewards and therefore use high discount factors in their Q-Value calculation. We investigate the behaviour of an adaptation of the VPG training algorithm on our task. We show that this adaptation can not accurately predict the required long-term action sequences. In addition to the limitations identified by Ewerton et al., it suffers from the known Deep Q-Learning problem of overestimated Q-Values. In an effort to solve our task, we turn to the Hourglass models and combine them with the Double Q-Learning approach. We show that this approach enables the models to accurately predict long-term action sequences when trained with large discount factors. Our results show that the Double Q-Learning technique is essential for training with very high discount factors, as the models Q-Value predictions diverge otherwise. We also experiment with different approaches for discount factor scheduling, loss calculation and exploration procedures. Our results show that the latter factors do not visibly influence the model's performance for our task.