Reinforcement Learning

Boost multi-task reinforcement learning with Cross-Task Policy Guidance (CTPG)! CTPG cleverly uses policies from already mastered tasks to guide the learning of new tasks, significantly improving effi…

This paper proposes a novel, statistically efficient offline policy evaluation method robust to environmental shifts and unobserved confounding, providing sharp bounds with theoretical guarantees.

SI2E, a novel RL exploration framework, leverages structural information principles to maximize value-conditional structural entropy, significantly outperforming state-of-the-art baselines in various …

EASI: Evolutionary Adversarial Simulator Identification bridges the reality gap in robotics by using GAN and ES to find optimal simulator parameters, enabling seamless sim-to-real transfer with minima…

DynaMITE-RL: A new meta-RL approach masters environments with evolving latent states by cleverly modeling episode sessions and refining existing meta-RL techniques.

Researchers developed a novel stochastic-process approach to precisely analyze learning dynamics in nonlinear perceptrons, revealing how input noise and learning rules significantly impact learning sp…

Dynamic Model Predictive Shielding (DMPS) ensures provably safe reinforcement learning by dynamically optimizing reinforcement learning objectives while maintaining provable safety, achieving higher r…

DCRL, a Dual Critic Reinforcement Learning framework, effectively mitigates high variance in reinforcement learning under partial observability by synergistically combining an oracle critic (with full…

Doubly Mild Generalization (DMG) improves offline reinforcement learning by selectively leveraging generalization beyond training data, achieving state-of-the-art results.

DoDont, a novel algorithm, uses instruction videos to guide unsupervised skill discovery, effectively learning desirable behaviors while avoiding undesirable ones in complex continuous control tasks.

Provably sample-efficient robust RL via interactive data collection is achieved by introducing the vanishing minimal value assumption to mitigate the curse of support shift, enabling near-optimal algo…

DiSPOs: a novel model for zero-shot policy optimization in reinforcement learning, enabling quick adaptation to new tasks by learning a distribution of successor features and avoiding compounding erro…

Sinkhorn distributional RL (SinkhornDRL) uses a regularized Wasserstein loss to improve distributional reinforcement learning.

DUSDi: A novel method for learning disentangled skills in unsupervised reinforcement learning, enabling efficient reuse for diverse downstream tasks.

DUPLEX: a novel RL method trains diverse, near-optimal policies in complex, dynamic environments by explicitly maximizing policy diversity using successor features. It outperforms existing methods in…

Diffusion-Reward Adversarial Imitation Learning (DRAIL) enhances Generative Adversarial Imitation Learning by integrating diffusion models, resulting in more stable and smoother reward functions for s…

Diffusion-DICE: A novel offline RL method using in-sample diffusion guidance for optimal policy transformation, achieving state-of-the-art performance.

QVPO, a novel online RL algorithm, leverages diffusion models’ multimodality to boost performance in continuous control tasks, overcoming limitations of unimodal policies.

DiCuRL uses diffusion models to generate challenging yet achievable RL training curricula, outperforming nine state-of-the-art methods.

Diffusion Spectral Representation (Diff-SR) enables efficient reinforcement learning by extracting sufficient value function representations from diffusion models, bypassing slow sampling and facilita…