One  of  the  most  crucial  challenges  faced  by  the  Li-ion  battery community  concerns  the  search  for  the minimum  time  charging  without  damaging  the  cells.   This  goal  can  be  achieved  by  solving  a  large-scale constrained optimal control problem which relies on accurate electrochemical models.  However, these models  are  limited  by  their  high  computational  cost  as  well  as  identifiability  and  observability  issues.   As  an alternative, simple output-feedback algorithms can be employed, but their performance strictly depends on trial  and  error  tuning.   Moreover,  particular  techniques  have  to  be  adopted  to  handle  safety  constraints.With  the  aim  of  overcoming  these  limitations,  we  propose  an  optimal-charging  procedure  based  on  deep reinforcement learning.  In particular, we focus on a policy gradient method to cope with continuous sets of states and actions.  First, we assume full state measurements from the Doyle-Fuller-Newman (DFN) model, which is projected to a lower-dimensional feature space via Principal Component Analysis.  Subsequently, this assumption is removed and only output measurements are considered as the agent observations.  Finally,  we  show  the adaptability  of  the proposed  policy  to changes in  the  environment’s  parameters.  The results are compared with other methodologies presented in the literature, such as the reference governor and proportional-integral-derivative approach.