The paper investigates the discrepancy between strong agents and their underlying weak world models in Atari Pong by reproducing five visual world-model agents and evaluating their frozen models. Closed‑loop rollouts reveal visual and dynamical failures such as ball disappearance and incorrect motion, while zero‑shot model‑based RL policies trained entirely within the frozen models perform poorly compared to the original agents. To address these issues, the authors introduce Concept‑Guided Spatial Regularization (CGSReg), an auxiliary loss focused on task‑critical ball regions, which improves both pixel‑space zero‑shot MBRL performance and closed‑loop rollouts for several agents.
arXiv:2607. 15142v1 Announce Type: new Abstract: World models are usually evaluated as components of model-based reinforcement learning (MBRL) systems, while the world models themselves are rarely studied in isolation.
arXiv:2602. 13977v2 Announce Type: replace-cross Abstract: Reinforcement learning (RL) promises to unlock capabilities beyond imitation learning for Vision--Language--Action (VLA) models, but its requirement for massive real-world interaction prevents direct deployment on physical robots.
arXiv:2410.14606v3 Announce Type: replace Abstract: Learning from a stream of experience as it arrives, also known as streaming learning, is a core part of natural learning. However, reliable streami...
The paper introduces Action-Contrastive Masked Transition Modeling (AC‑MTM), a method that stabilizes Joint‑Embedding Predictive Architectures (JEPAs) without relying on Gaussian regularization. AC‑MTM adds a training‑only inverse‑dynamics head that uses Action‑NCE to force each latent transition to identify its generating action, thereby preventing encoder collapse. Experiments on pixel‑control and multi‑object visual tasks show that AC‑MTM trains stably from scratch and matches or surpasses the performance of SIGReg, achieving up to a 24‑point improvement on the OGBench Visual Scene benchmark.
arXiv:2607. 02431v1 Announce Type: cross Abstract: Reinforcement learning (RL) can overcome the demonstration-coverage limitation of imitation learning (IL) by allowing robots to improve through trial-and-error interaction beyond the states observed in demonstrations.
The paper introduces Action-Contrastive Masked Transition Modeling (AC‑MTM), a method that replaces the Gaussian regularizer used in Joint‑Embedding Predictive Architectures (JEPAs) with a contrastive inverse‑dynamics head. AC‑MTM trains a forward latent‑prediction model while an auxiliary inverse‑dynamics task forces the encoder to distinguish actions from latent transitions, preventing collapse without requiring a target network or reconstruction loss. Experiments on pixel‑control and multi‑object visual tasks show that AC‑MTM matches or surpasses the performance of the Gaussian‑based SIGReg regularizer, achieving up to 20–24 point improvements on the OGBench Visual Scene benchmark.
PAVXploreRL introduces a reinforcement learning framework that builds on a pretrained latent world model to explicitly optimize Physical Plausibility, Action Adherence, and Visual Fidelity (PAV) objectives. By combining in‑distribution expert trajectories with noise‑driven out‑of‑distribution action exploration, the method avoids reliance on paired video supervision and improves generalization. Experiments demonstrate a 5.6% average performance gain over pretrained baselines and more reliable policy evaluation with reduced overestimation bias.
arXiv:2606. 12200v1 Announce Type: cross Abstract: We study policy representation learning from unlabeled multi-policy behavioral data.
arXiv:2607. 01531v2 Announce Type: replace Abstract: Learning how an environment behaves from interaction is central to building agents that adapt to unfamiliar tasks.
CLAP is a cross-embodiment framework for action‑conditioned video generation that can be trained on diverse internet‑scale videos from both humans and robots. It reconciles different action spaces—end‑effector poses, language instructions, and latent actions—using a curriculum that first learns physics priors from unlabeled video and then grounds them in real‑world action spaces for zero‑shot deployment. The resulting models match or exceed state‑of‑the‑art single‑embodiment models in challenging environments and support few‑shot adaptation across a wide range of robot morphologies.
arXiv:2609.37907v1 Announce Type: new Abstract: Video games offer scalable environments for studying perception and control in embodied agents.Abundant online gameplay videos could supply demonstrati...
The paper presents Real‑Time EXPO‑FT, a reinforcement learning framework that fine‑tunes large Vision‑Language‑Action models for real‑time robotic control. It separates slow, expressive action generation from fast, reactive edits, allowing a lightweight policy to adjust actions based on the latest observation. Experiments on the Kinetix benchmark and four dynamic real‑world tasks show that Real‑Time EXPO‑FT achieves superior performance, improving policy success rates from 42% to 97% with only ten minutes of online data and no human intervention.