Simulation to Real-World Transfer allows affordable and fast training of learning-based robots for manipulation tasks using Deep Reinforcement Learning methods. Currently, Sim2Real uses Asymmetric Actor-Critic approaches to reduce the rich idealized features in simulation to the accessible ones in the real world. However, the feature reduction from the simulation to the real world is conducted through an empirically defined one-step curtail. Small feature reduction does not sufficiently remove the actor's features, which may still cause difficulty setting up the physical system, while large feature reduction may cause difficulty and inefficiency in training. To address this issue, we proposed Curriculum-based Sensing Reduction to enable the actor to start with the same rich feature space as the critic and then get rid of the hard-to-extract features step-by-step for higher training performance and better adaptation for real-world feature space. The reduced features are replaced with random signals from a Deep Random Generator to remove the dependency between the output and the removed features and avoid creating new dependencies. The methods are evaluated on the Allegro robot hand in a real-world in-hand manipulation task. The results show that our methods have faster training and higher task performance than baselines and can solve real-world tasks when selected tactile features are reduced.
I. INTRODUCTIONDexterous in-hand manipulation is one of the essential functions for robots in human-robot interaction [1], intelligent manufacturing [2], telemanipulation [3], and assisted living [4], but it is also hard to solve due to the high degrees of freedom (DoFs) in control space and the complex interaction with the object. Deep Reinforcement Learning (DRL) [5] has shown its abilities in recent research [6][7][8] to solve dexterous in-hand manipulation tasks thanks to its learning capability, which enables the robot to find a control policy by interacting with the environment through exploration and exploitation.Recent literature uses Simulation to Real-world (Sim2Real) [9] transfer, which trains the DRL policy in the simulated environment and then transfers the policy to the real robot to complete the same task. The benefit of using Sim2Real is that the simulation platform can be easily customized to recreate the real-world environment, reducing the implementation effort. The training can be accelerated with multi-thread and parallel training setups [6]. Most importantly, the simulation environment can provide more explicit information [10] that is hard to extract in the real world, such as tactile, depth, and
