Digest

Reset

1180 papers

2026-10-01 · arXiv:2609.39751

Beyond Policy Alignment: Closing the Planning-Learning Loop for Robot Control with Learned World Models

Kowndinya Boyalakuntla, Yuhan Liu, Abdeslam Boularias

The authors propose PL-MPC, an extension of policy-constrained TD-M(PC)2 that modifies (1) critic supervision via hybrid multi-step TD targets, (2) MPC terminal-value evaluation via disagreement-aware terminal estimates, and (3) actor learning via return-weighted actor distillation from planner-executed actions. They report largest gains on HumanoidBench balance-hard and hurdle, and higher zero-shot sim-to-real success on wrench–nut alignment.

DrivingAI summary
2026-10-01 · arXiv:2609.39763

DiffWAM: A Fast and Efficient Navigation World Action Model

Mo Zhu, Yuze Wu, Xijie Huang et al.

DiffWAM is a geometry-conditioned navigation world-action model that decodes continuous UAV camera trajectories directly from intermediate predictive features of a frozen video foundation model. It uses Grid-Motion and Latent2Pose with first-frame geometry, and needs complete future-video rollout + geometric reconstruction only for offline supervision. FastDreamer further overlaps predictive/geometry computation with flight-time execution and uses timestamp-aware trajectory handoff.

DrivingAI summary
2026-10-01 · arXiv:2609.39870

Magic-W0: A Structured World-Action Foundation Model for Physical Intelligence

Xuhua Chen, Zhenhan Yin, Yuan Zhang et al.

Magic-W0 is a world-action foundation model that organizes robot interaction as a Structured World Transition: Current State–Transition–Future State. Current State uses VLM context plus Current 3D Geometry, Transition is 3D Motion induced by actions, and Future State uses Future Semantics. A layer-aligned bidirectional architecture couples action hypotheses with structured world prediction to inform continuous action generation.

RoboticsAI summary
2026-10-01 · arXiv:2609.39873

SplineWAM: Adaptive Action Horizons for World Action Models via B-Spline Representations

Jun Guo, Xiaoshen Han, Qiwei Li et al.

SplineWAM replaces fixed-length action chunking in world action models with a fixed-size window of adaptive cubic B-spline parameters. The predicted knot times determine the decoded action duration and the interval until the next policy call. For asynchronous deployment, JP-RTC enforces continuity on decoded actions via a decoder Jacobian pullback. The authors report higher success with fewer policy calls on LIBERO-Plus and RoboCasa, and real-robot gains/match on three bimanual tasks under asynchronous execution.

DrivingAI summary
2026-10-01 · arXiv:2609.39970

PhasePlan: Ordered Future-Phase Planning for Robot Brain Models

Xiaoyu Yang, Yafei Zhang, Wensheng Li et al.

PhasePlan is an ordered future-phase planning method for robot brain models. From current multimodal observations, it predicts a task phase distribution at each future action position and uses the per-position plan representations to condition the corresponding actions. The planner is trained first and then frozen during action-model adaptation to preserve stable phase-transition structure. Results on conveyor-belt manipulation reduce offline joint-action MAE by ~22.5% vs the original π0.5 model.

RoboticsAI summary
2026-10-01 · arXiv:2609.39973

EWAM: Emergent Depth-Wise Specialization in a Unified Embodied Model -- From Semantic Understanding through Visual Foresight to Action

Hao Wang, Jiajun Wen, Jingzhi Liu et al.

EWAM is an action-centric unified embodied model with three layer-aligned experts (vision-language, video, action) and asymmetric joint attention where action tokens read semantic, current-visual, predicted-future, and action information at every layer. Without layer-wise supervision, it develops emergent depth-wise specialization: shallow layers emphasize vision-language, intermediate layers predicted futures, and deep layers action self-attention. The authors report superior performance in simulation and real robots.

RoboticsAI summary
2026-10-01 · arXiv:2609.40003

DashVMC: Real-Time Discrete World Model Control in Geometry Dash

Florent Tariolle, Florian Yger

DashVMC learns a compact, action-conditioned discrete world model from ~2 hours of recorded Geometry Dash gameplay, then improves a BC-initialized controller using PPO entirely in frozen-model rollouts (no live-game interaction). The refined policies survive longer than BC across three controller seeds on all three official levels and a held-out community layout. A deployment baseline skips visual generation and sustains a 60-Hz capture-to-action loop on an RTX 2060.

DrivingAI summary
2026-10-01 · arXiv:2609.40129

VR-JEPA: Learning Contrastive-State Latent Guidance for Generation-based Video Reasoning

Zehua Ma, Kun Xiang, Yunshuang Nie et al.

VR-JEPA uses V-JEPA latent predictions to guide generation-based video reasoning. It trains a latent reasoner with localized contrastive-state learning by comparing successful reference trajectories to task-matched generated candidates, and uses a skill-routed mixture of experts (SR-MoE) to specialize across reasoning skills. The predicted latent trajectories condition a video diffusion generator to improve task consistency.

2026-10-01 · arXiv:2609.40153

Dream4ACT: A Shared Visual Action Interface for Multi-Embodiment Video-Action Modeling

Xiangyu Zhu, Jin Xu, Yue Guo et al.

Dream4ACT is a video–action world model for multi-embodiment joint control. It unifies heterogeneous joint spaces using “action views”: URDF-based forward-kinematics renderings of target joint configurations from four fixed virtual cameras. A shared VAE+diffusion transformer models RGB observations and four action-view streams with masked flow matching for forward dynamics, inverse dynamics, and joint generation; predicted views are converted to executable joints via training-free URDF-constrained multiview recovery.

RoboticsAI summary
2026-10-01 · arXiv:2609.40177

Social-WM: Safety-Aware Latent World Models for Robot Social Navigation

Zhihao Zheng, Mooi Choo Chuah

Social-WM is an action-conditioned latent world-model planning framework for robot social navigation from egocentric RGB video. It distinguishes nominal actions from realizable actions using safety-aware future prediction (targeting the actual observed future) plus a realizable inverse-dynamics objective. During planning, it imagines action chunks and uses nominal–realizable discrepancy as a safety signal to select safer candidates.

Latent DynamicsAI summary
2026-10-01 · arXiv:2609.40219

Learning Skills from Historical Action Trajectories: Action Experience Dictionary for World Action Models

Qi Lyu, Jiahua Dong, Hao Shen et al.

The authors report an Action Experience Dictionary (AED) for World Action Models that encodes historical action trajectories into shared action embeddings for cross-task skill reuse. They retrieve AED embeddings with a pretrained action tokenizer, visually condition them via cross-attention, and prepend them to action tokens for prediction. A motion-aware transition loss supervises visual feature change over random temporal intervals to reduce reliance on background cues. They evaluate on LIBERO, RoboTwin, and real-world cross-embodiment tasks.

RoboticsAI summary
2026-10-01 · arXiv:2609.40222

LOCI: Spatial Linear Memory for Streaming World Models

Ji Xia, Tingting Liao, Xuezhi Liang et al.

LOCI is a streaming video world model for revisits. It uses a hybrid spatial memory: half the transformer blocks keep a key–value cache of past observations, while the other half uses a projectively conditioned recurrent linear-attention memory with camera-geometry-aware reads/writes. Recurrent readouts shape subsequent historical queries. The authors report better revisit fidelity than representative world models and a same-recipe full-softmax, and lower peak memory with full history (~30%).

2026-10-01 · arXiv:2609.40341

Ego4WAM: What Matters When Scaling Egocentric Human Data for Robot Learning?

Zhihao Sun, Liu Liu, Xinjiang Wang et al.

Ego4WAM studies how egocentric human data properties affect downstream robot learning under a fixed world-action model framework. The authors show that aligned human demonstrations improve out-of-distribution generalization and reduce target-task robot data needs; data duration and task diversity affect different capabilities differently; and video-only supervision remains effective without action labels, helping subsequent video-action training. Closed-loop evaluation is done on real robots and RoboDojo.

DrivingAI summary
2026-10-01 · arXiv:2609.40358

Physis-Lang: Self-Evolving Language as a Physical Representation for Video World Model

Liming Lu, Xianzheng Ma, Wenkun He et al.

Physis-Lang is a self-evolving language framework for video world models. It represents physical processes with explicit “physics_reasoning” and uses an agentic caption-evolution loop plus PhysCapBench (precision/recall on atomic physical assertions) to iteratively refine caption instructions. It also converts model failures into deficiency profiles for language-guided retrieval and re-captioning, then fine-tunes video backbones (Wan/Cosmos) with the evolved physical captions, improving physical plausibility and surpassing Veo 3.1 starting from Cosmos3-Nano.

2026-09-30 · arXiv:2609.31657

Enhancing Foundation Models for Imbalanced SAR Ship Classification via Targeted Oversampling

Ch Muhammad Awais, Marco Reggiannini, Davide Moroni

The authors benchmark frozen DOFA and SAR-JEPA embeddings on imbalanced OpenSARShip and mitigate long-tail imbalance without backbone fine-tuning by applying minority-only oversampling in embedding space, then training a lightweight classifier head. Across both foundation models, oversampling improves Macro-F1 and accuracy vs baselines, with the largest Macro-F1 gains: DOFA+ADASYN (34.39→38.56) and SAR-JEPA+SVM-SMOTE (25.89→32.30).

JEPAAI summary
2026-09-30 · arXiv:2609.31680

Does Joint-Embedding Predictive Architecture Pretraining Help Time Series Forecasting?

Yutong Feng, Bowen Liao, See Kiong Ng et al.

The authors evaluate one JEPA instantiation for time-series forecasting across nine backbones and eleven benchmarks. They report that JEPA’s benefit is highly inconsistent across backbones: it gives consistent gains for some architectures and consistent degradation for others, even on the same dataset. The authors find the variability holds across both temporal and spatio-temporal task families.

JEPAAI summary
2026-09-30 · arXiv:2609.31893

CyberWorld: World Models for Sample-Efficient Autonomous Cyber Defense

Ryozo Masukawa, Sanggeon Yun, Raheeb Hassan et al.

CyberWorld is a Dreamer-style world model for autonomous cyber defense that learns latent cyber dynamics from vector, graph, textual, and multimodal representations of a defended network. Using CyberWheel, the graph-based variant reaches the deploy_then_stop control after 3.6k–15.8k environment steps (vs model-free PPO needing 2.3M–3.1M steps or failing within a 3.2M budget).

DrivingAI summary
2026-09-30 · arXiv:2609.31938

Cache-Aware Conv3D Lowering Across Embedded World-Model Decoders

Jiaming Zhang, Wu Yang, Shuai Tao et al.

Cache-aware Conv3D lowering for embedded generative video VAEs: supported causal Conv3D calls are expressed as batched spatial Conv2D while preserving pretrained weights, causal-cache semantics, convolution parameters, bias placement, and output layout. On 64-GB Jetson AGX Orin in the Cosmos3-Edge image-to-video pipeline, the authors report ~7.32× VAE-decoder speedup and 2.21× end-to-end speedup at 25 frames.

2026-09-30 · arXiv:2609.32268

Self-Reconstruction Dynamics for Autoencoder Reconstruction Refinement

Hitoshi Iyatomi

The authors study Self-Reconstruction Dynamics (SRD), produced by repeatedly applying a frozen autoencoder to its own reconstruction, which forms transient image/latent trajectories. Although repeated self-reconstruction degrades fidelity, SRD encodes sample-specific correction information. They propose SRD-guided Reconstruction Refinement (SRD-RR), predicting a latent correction from a short SRD with the AE frozen and no per-sample test-time optimization. Across six datasets, SRD-RR recovers 38.6% (one) to 45.3% (two) of the empirically recoverable MSE gap.

Latent DynamicsAI summary
2026-09-30 · arXiv:2609.32322

Not All Errors Matter: Decision-Relevant Prediction Error Predicts Planning Quality

Linhao Wang, Yiyan Fan, Dongjin Huang

The authors show that two world models with similar total prediction error can yield very different planning outcomes when their errors occur on different state dimensions. They introduce Decision-Relevant Prediction Error (DRPE), measuring multi-step error only on decision-relevant state dimensions, and an iso-error protocol that varies error allocation while keeping total error fixed.

Latent DynamicsAI summary