Raghuram Bharadwaj Diddigi

Language instructions can replace costly trajectory-based updates in meta-reinforcement learning, leading to faster training without sacrificing performance.

Full gradient updates can stabilize successor feature learning, unlocking more effective transfer in RL.