arXiv:2607. 19749v1 Announce Type: cross Abstract: Model-based reinforcement-learning agents of the DreamerV3 family forget catastrophically when trained on task sequences, even when an unbounded replay buffer preserves every earlier experience.
The paper introduces a compositional continual learning benchmark for world models in robot manipulation, designed to isolate knowledge reuse from learning speed and capacity. Tasks are curated to combine previously seen action and perception components, allowing analysis of how different modalities affect reuse. Experiments show that modular world models better balance reuse and forgetting than conventional methods, yet none fully solve the challenge, highlighting the need for models explicitly built to reuse knowledge without forgetting.
RIFAR is a new continual learning method for robots that uses reliability screening and drift-aware replay to mitigate forgetting while keeping storage low. It reconstructs past trajectories from short demonstration prefixes and employs a frozen inverse-dynamics model to verify action‑visual consistency. In experiments on LIBERO suites and real‑world tasks, RIFAR outperforms previous generative replay approaches, achieving high performance with only a small fraction of stored steps.
arXiv:2609.39235v1 Announce Type: cross Abstract: World models offer a promising way to help robots understand how the physical world evolves and plan complex behaviours through imagination. Yet exis...
arXiv:2607.09042v2 Announce Type: replace Abstract: Reinforcement learning is increasingly used to fine-tune vision-language-action (VLA) models, but robot interaction is expensive and learning becom...
The paper proposes a method to distill world‑model representations into compact Vision‑Language‑Action (VLA) policies. By adding a single feature‑alignment term during VLA training, a frozen world model’s internal features are cached and the student policy learns to match them, eliminating the need for a generative future‑rolling component. The resulting lightweight policy runs in 32 ms on an RTX 5090, achieving high performance on LIBERO and RoboCasa‑GR1, and transfers effectively to real robotic hardware.
MemBodied introduces a fixed‑size episodic memory for Vision‑Language‑Action models, comprising an associative state that tracks interactions across policy calls and an episode anchor that stores a compact representation of the initial scene. By conditioning action generation on these memory components instead of raw past observations, MemBodied reduces context bloat and inference latency. In five memory‑dependent RMBench tasks, it outperforms stateless and vanilla recurrent policies by significant margins, and achieves a 90.6% success rate on the LIBERO‑Long suite, improving over the baseline by 5.4%.
arXiv:2607. 09042v1 Announce Type: new Abstract: Reinforcement learning (RL) is increasingly used to post-train vision-language-action (VLA) models, but every update consumes robot rollouts that are slow and costly to collect, making sample efficiency a central concern.
arXiv:2603. 24576v2 Announce Type: replace-cross Abstract: Robots often observe information that determines a future action long before that action is executed.
The paper introduces Divide-and-Remember (D&R), a recursive memory method for vision-language-action (VLA) policies that optimises memory by maximizing the conditional mutual information between actions and memory given observations. D&R recursively divides the full history into top‑K selections over 2K tokens, using a shared lightweight selector across all recursion blocks to handle unbounded histories efficiently. Evaluated on the RoboMME benchmark of 16 long‑horizon manipulation tasks, D&R achieves state‑of‑the‑art success rates with consistent gains across all suites while using only 64 tokens, and similar improvements are observed in real‑robot experiments.
arXiv:2608. 03483v1 Announce Type: cross Abstract: Existing chunk-based Vision-Language-Action (VLA) models execute a fixed number of actions (i.
FIRM-WM is a compact pixel world model that separates a goal‑comparable configuration from a 128‑dimensional dynamic fiber, enabling reward‑free visual planning from offline videos. It addresses two key mismatches: aligning planning states with goal images and reconciling factual trajectories with interventional sampling. In experiments, FIRM‑WM achieves high success rates on TwoRoom, Reacher, and OGBench‑Cube while using fewer parameters and faster planning times than prior models.