We present Coarse-to-fine Q-Network with Action Sequence (CQN-AS), a value-based RL algorithm that trains a critic network to output Q-values over a sequence of actions.

We present Coarse-to-fine Reinforcement Learning (CRL), which trains RL agents to zoom-into continuous action space in a coarse-to-fine manner. Within this framework, we present Coarse-to-fine Q-Network (CQN), a value-based RL algorithm for continuous control.

We present BiGym, a new benchmark and learning environment for mobile bi-manual demo-driven robotic manipulation. BiGym consists of 40 diverse tasks in home environment, and provides human-collected demonstrations.

We propose Masked Multimodal Learning (M3L), which jointly learns a policy and visual-tactile representations based on masked autoencoding.

We present RSP, a framework for visual representation learning from videos, that learns representations that capture temporal information between frames by training a stochastic future frame prediction model.

We introduce Render and Diffuse (R&D), a method that unifies low-level robot actions and RGB observations within the image space using virtual renders of the 3D model of the robot.

We introduce a new exploration technique that maximizes value-conditional state entropy, which takes into account the value estimates of states for computing the intrinsic bonus.

We present ARP (Adaptive Return-conditioned Policy), which utilizes an adaptive multimodal reward signal for behavior learning.

We introduce MV-MWM that learns multi-view representations via masked view reconstruction and utilize them for visual robotic manipulation.

We introduce LAMP, a method for pretraining RL agents by using multimodal reward signal from Video-Langauge models.

We introduce PIG, a simple yet effective self-imitation scheme which distills a subgoal-conditioned policy into the target-goal-conditioned policy.

We introduce MWM that learns a latent dynamics model on top of an autoencoder trained with convolutional feature masking and reward prediction.

We introduce DADT, which improves dynamics generalization of decision transformer by introducing a next-state prediction objective.

We introduce APV that can leverage diverse videos from different domains for pre-training to improve sample-efficiency.

We introduce a video prediction model that can generate 256x256 frames by training an autorgressive transformer on top of VQ-GAN.

We introduce semi-supervised learning and temporal data augmentation for improving the feedback-efficiency of preferenced-based RL.

We introduce OREO, a regularization technique for behavior cloning from pixels.

We introduce HIGL, a goal-conditioned hierarchical RL method that samples landmarks and utilizes them for guiding the training of a high-level policy.

We introduce an offline-to-online RL algorithm to address the distribution shift that arises during the transition between offline RL and online RL.

We introduce RE3, an exploration technique for visual RL that utilizes a k-NN state entropy estimate in the representation space of a randomly initialized & fixed CNN encoder.

We introduce T-MCL that learns multi-headed dynamics model whose each prediction head is specialized in certain environments with similar dynamics, i.e., clustering environments.

We introduce CaDM that learns a context encoder to extract contextual information from recent history with self-supervised losses of future and backward prediction.

We introduce LwD, a deep RL framework for the maximum independent set problem.