Assistive robots introduce a new paradigm for developing advanced personalized services. At the same time, the variability and stochasticity of environments, hardware and unknown parameters of the interaction complicates their modelling, as in the case of staircase traversal. For this task, we propose to treat the problem of robot configuration control within a reinforcement learning framework, using policy gradient optimization. In particular, we examine the use of safety or traction measures as a means for endowing the learned policy with desired properties. Using the proposed framework, we present extensive qualitative and quantitative results where a simulated robot learns to negotiate staircases of variable size, while being subjected to different levels of sensing noise.

Autonomous control of reconfigurable robots is crucial for their deployment in diverse environments. The development of such skills is however hampered by the diversity in hardware and task constraints. We advocate the use of artificial intelligence-based approaches to improve scalability to different conditions and portability to platforms of comparable traversability skills. In particular, we succeed in tackling the problem of staircase traversal via a reinforcement learning-based control framework applicable to different articulated tracked robots, powerful enough to generalize to varying conditions learnt in simulation and to transfer to reality in a zero-shot setting. Our extensive experiments demonstrate the robustness of the framework in learning tasks with increased risk and difficulty induced by platform diversification and increased control dimensionality.

A simulation framework based on the open-source robotic software Gazebo and the Robot Operating System (ROS) is presented for articulated tracked robots, designed for reinforcement learning-based (RL) control skill acquisition. In particular, it is destined to serve as a research tool in the development and evaluation of methods in the domain of mobility learning for articulated tracked robots, in 3D indoor environments. Its architecture allows to interchange between different RL libraries and algorithm implementations, while learning can be customized to endow specific properties within a control skill. To demonstrate its utility, we focus on the most demanding case of staircase ascent and descent using depth image data, while respecting safety via reward function shaping and incremental, domain randomization-based, end-to-end learning.