From Trajectories to Instructions: Language-Conditioned Meta-Reinforcement Learning

This paper introduces LA-MAML, a language-adapted variant of Model-Agnostic Meta-Learning for reinforcement learning. Instead of collecting task trajectories and performing gradient-based inner-loop updates, LA-MAML uses a learned embedding of the task instruction to adapt global policy parameters in a single step. On the BabyAI benchmark, the authors report competitive or improved performance over baselines while substantially reducing per-iteration wall-clock training time. The result suggests that language instructions can provide an efficient task-specific adaptation signal for meta-RL when task descriptions are available.