Language for Model-Based Reinforcement Learning
Multi-task natural-language control of a robotic arm via seq2seq trajectory generation.
Applied multi-task learning to robotic manipulation by using natural language commands to control a robotic arm. A seq2seq architecture, incorporating autoencoders and attention, translates natural commands into trajectories for the arm. This approach demonstrated improved performance and generalization on manipulation tasks compared to using model-based reinforcement learning alone.
