A new test-time framework called World-Model Policy Arbiter (WMPA) raises the macro-average success rate on 18 state-based OGBench datasets from 44% to 58% by rolling out frozen goal-conditioned polic…
MIT researchers introduced LASER, an offline reinforcement learning algorithm that applies latent-space adjoint matching to achieve entropy-regularized latent-space RL with expressive flow policies wh…
Researchers proposed Reward Stimulation Implicit Q-Learning (RSIQL), a non-hierarchical offline goal-conditioned reinforcement learning method that adds reward signals at progress-making intermediate …
Researchers introduced a noisy-space action-value (Q-)function and a noisy-space policy gradient (NSPG) that trains diffusion policies for offline reinforcement learning without backpropagating throug…
Researchers propose CoDrift, a compositional framework for offline reinforcement learning that combines three objective-level fields into a unified policy field, achieving the best average rank across…
Researchers propose Hierarchical Implicit Q-Chunking (HiQC), an offline goal-conditioned reinforcement learning algorithm that combines high-level latent planning with low-level action chunking to add…
The Agency for Defense Development (ADD) introduced Diffusion ReRoll, a method that selectively re-noises resolved regions during sequential denoising to enable cross-horizon revision in robotic seque…
DAGR, a new approach to goal-conditioned reinforcement learning, refines state-independent goal encoders into state-conditioned ones using multi-scale gated cross-attention, improving navigation tasks…
Researchers have identified that the real cause of training instability in reinforcement learning with flow-matching policies is not iterative action generation but the vanilla sampling strategy, and …
Researchers introduced Latent Policy Steering (LPS), a new offline reinforcement learning method that directly integrates Q-gradients from the original action space into latent-space optimization, ach…
Researchers have developed CARL (Contrastive Action-based Representations for Reusable Local Control), a new algorithm that exploits local dynamics regularity to learn reusable skills for hierarchical…