We introduce a novel model-based offline RL method, Lower Expectile Q-learning (LEQ), which enhances long-horizon task performance by mitigating the high bias in model-based value estimation via expectile regression of λ-returns. Our empirical results show that LEQ significantly outperforms previous model-based offline RL methods on long-horizon tasks, such as the D4RL AntMaze tasks, matching or surpassing the performance of model-free approaches.

We propose a novel unsupervised goal-conditioned RL method, TLDR, which leverages TemporaL Distance-aware Representations. Our approach selects faraway goals to initiate exploration and compute intrinsic exploration rewards and goal-reaching rewards. Our experimental results in robotic locomotion and manipulation environments demonstrate that our method significantly outperforms previous unsupervised GCRL methods in achieving a wide variety of states.

We present VECA(Virtual Environment for Cognitive Assessment), which consists of two main components: (i) a first benchmark to assess the overall cognitive development of an AI agent, and (ii) a novel toolkit to generate diverse and distinct cognitive tasks. VECA benchmark virtually implements the cognitive scale of Bayley Scales of Infant and Toddler Development-IV(Bayley-4), the gold-standard developmental assessment for human infants and toddlers.

We investigate the emergence of critical periods on multimodal reinforcement learning. We show the performance on RL task and transfer learning depends on what and when the guidance is given to the agent.

Toddler's learning procedure consists of interactive experiences, resulting in task-agnostic representations. Inspired by those precedures, we pretrain the agent on a visual navigation task and show that the representations obtained during the RL task is expandable to various vision tasks.