• soft robot
• imitation learning
• behavioral cloning
• domain adaptation
• BCO

Soft robotics is a promising class of robotics, dealing with compliant nature-inspired robots. It introduces substantial potential for applications in a wide range of fields, even if significant challenges in modelling and control appear.
For this reason, it is interesting to explore alternative ways to make these kind of agents learn.
Rather than manually engineer a desired behavior, it is possible to achieve it through the process of learning from demonstrations of a teacher.
The aim of this research is to elicit in a pneumatically actuated soft robot, the movement observed from a human arm. The stochasticity of soft robotic movements, unavailability of the expert actions and need of domain remapping are some of the challenges encountered.
Behavioral Cloning from observation is the approach employed, meaning that the policy able to reproduce autonomously the shown behavior is learnt as direct mapping from states to control input, and that state-only demonstrations are provided.
In first phase, the robot explores the workspace to map the relationship between the input pressures signals and the output positions in the task space.
Secondly, for each type of movement, a policy model is learnt after the robot has executed the pre-processed and remapped demonstration using the actions suggested by the inverse dynamic model.
Since the main issues in the algorithm were related to hardware limitations, future work should start from the adoption of more suitable instrumentation.