The process of learning a manipulation task depends strongly on the action\nspace used for exploration: posed in the incorrect action space, solving a task\nwith reinforcement learning can be drastically inefficient. Additionally,\nsimilar tasks or instances of the same task family impose latent manifold\nconstraints on the most effective action space: the task family can be best\nsolved with actions in a manifold of the entire action space of the robot.\nCombining these insights we present LASER, a method to learn latent action\nspaces for efficient reinforcement learning. LASER factorizes the learning\nproblem into two sub-problems, namely action space learning and policy learning\nin the new action space. It leverages data from similar manipulation task\ninstances, either from an offline expert or online during policy learning, and\nlearns from these trajectories a mapping from the original to a latent action\nspace. LASER is trained as a variational encoder-decoder model to map raw\nactions into a disentangled latent action space while maintaining action\nreconstruction and latent space dynamic consistency. We evaluate LASER on two\ncontact-rich robotic tasks in simulation, and analyze the benefit of policy\nlearning in the generated latent action space. We show improved sample\nefficiency compared to the original action space from better alignment of the\naction space to the task space, as we observe with visualizations of the\nlearned action space manifold. Additional details:\nhttps://www.pair.toronto.edu/laser\n