Digest

Reset

1167 papers

2026-10-22 · arXiv:2610.29435

Events, Not Pixels: Where World Models Get the Return Wrong

Haoyue Liu, Xiaoying Tang

The authors analyze why pixel world models can mis-rank actions during planning: return error can concentrate at “event transitions” (reward earned or episode ends) even when those transitions look visually normal. They introduce a transition-level substitution diagnostic that swaps imagined transitions with real ones under budget-matched random controls, and show event patches remove much more return error than patches chosen by pixel metrics.

Latent DynamicsAI summary
2026-10-22 · arXiv:2610.29528

Are Joint-Embedding Predictive Objectives Effective for Longitudinal EHRs?

Ashvin Gupta, Brendan Delaney, Alessandra Russo

The authors adapt Joint-Embedding Predictive Architectures (JEPAs) to longitudinal EHR sequences (MIMIC-IV) using three target constructions (Span, Future, Random) and compare against masked language modelling (MLM). Although all JEPA variants train stably without representational collapse, the authors report that all three consistently underperform standard MLM on downstream clinical prediction tasks.

JEPAAI summary
2026-10-22 · arXiv:2610.29677

Toward Recursive Self-Improvement in Physical AI

Meng Luo, Zechen Bai, Shanqing Xu et al.

The authors review recursive self-improvement (RSI) in physical robot learning: systems that improve not only behavior but the process of improvement. They define when gains carry forward in four conditions tied to physical constraints (informative feedback, usability, robustness to distribution shift, continued practice) and propose recursion tests requiring retained repair knowledge. They argue existing physical studies support earlier improvement claims but haven’t isolated recursion.

RoboticsAI summary
2026-10-22 · arXiv:2610.29702

CausGround: Causal Interaction Grounding for Adaptive Embodied Planning

Yijun Shen, Diqi He, Jinbang Guo et al.

CausGround is a non-parametric causal interaction grounding framework for embodied planning. It extracts action preconditions and effects from experience into a factorized world-state representation (object attributes + inter-object relations). At planning time it grounds a task-relevant causal graph to enable backward reasoning over preconditions and forward reasoning over effects, improving planning and interaction reasoning across VirtualHome, RoboCasa, SWITCH, and real-robot manipulation without parameter updates.

DrivingAI summary
2026-10-22 · arXiv:2610.29829

Robot Generalization Beyond Volume: The Diversity Path via Zero-Shot Capabilities

Rafael Felix

The authors formulate zero-shot generalization in physical AI as coverage of the deployment distribution, arguing that increasing samples from the same collectors shrinks the unseen region only slowly, while each new contributor can extend it. They propose a shared latent space for robot actions combining multi-modal alignment, generative augmentation, and noisy-label learning, plus safety signals via seen/unseen detection and out-of-distribution noise estimation. They also propose a developer community building on Origami.

DrivingAI summary
2026-10-22 · arXiv:2610.29915

CausalAlign: Causal Alignment of Video Generation Models for Consistent Ego-Motion

Kai Zhu, Yintianrun Zhang, Shuwei Jiang et al.

CausalAlign diagnoses where ego-motion information is represented, used, and forgotten inside diffusion-based video generation models, then preserves it by aligning VGM representations to 3D delta-token supervision. Using Wan2.2 fine-tuned on DL3DV and RealEstate10K, the authors report improved motion accuracy, geometric consistency, and temporal smoothness, while maintaining visual quality. They also report gains for world action models on mobile manipulation.

DrivingAI summary
2026-10-22 · arXiv:2610.29989

Looking Ahead: Modeling Future Evolution in Autoregressive Video Generation

Ruopeng Gao, Yukun Huang, Jiarui Hu et al.

The authors identify a shortcut in teacher-forced autoregressive video diffusion: strong temporal redundancy lets local continuity help denoise the current chunk. They propose Future Chunk Prediction (FCP), an auxiliary training objective that predicts temporally separated future chunks without access to intervening visual context, requiring no extra pretrained visual encoder and leaving autoregressive inference unchanged. FCP improves physical plausibility with lower overall training compute budget in controlled comparisons.

Video World ModelsCode availableAI summary
2026-10-22 · arXiv:2610.30097

AtomicRL: Atomic Reinforcement Learning for Long-Horizon Robotic Manipulation

Likui Zhang, Zhe Zhu, Zhihao Zhan et al.

AtomicRL is an RL post-training framework for long-horizon robotic manipulation with vision-language-action policies. It uses SAPO to gate advantage/TD-residual propagation at atomic subtask semantic boundaries to reduce credit contamination from failed later subtasks, and BATE to sample tasks and adjust exploration based on subtask mastery. It also adds an automated Real2Sim2Real pipeline for real-robot transfer.

RoboticsAI summary
2026-10-22 · arXiv:2610.30117

CW-WAM: A Structured Contact-Wrench World Action Model for Force-Sensitive Manipulation

Houzhu Ding, Xinda Qi, Luqi Wang et al.

CW-WAM is a world action model for force-sensitive manipulation that jointly predicts video, motion, and contact wrench. It uses a unified interaction module with paired action–wrench history, pair-causal attention, and long-term wrench memory, and drives a 20-Hz outer-loop admittance controller in a slow–fast system. On real robots, it raises insertion success from 25% to 70% (peg) and from 40% to 95% (adapter) vs WAM without wrench input.

RoboticsAI summary
2026-10-22 · arXiv:2610.30120

ALPS: Visual State Encoders for World Models

Yanqing Liu, Letian Zhang, Chengzhi Liu et al.

ALPS is a pretrained ViT-B/16 visual state encoder that unifies physical-state information, native language alignment, and pixel reconstruction in a single set of patch tokens. It is pretrained from scratch using language, reconstruction, and dense spatial correspondence objectives on still images and text, then refined via cross-view self-distillation. Frozen ALPS tokens support low LIBERO state-probe error and strong LIBERO policy success, and enable decoding predicted tokens into images for inspection.

Latent DynamicsAI summary
2026-10-22 · arXiv:2610.30125

PhysiCalWorld: In-Context Physics Calibration for World Models

Yunhao Luo, Wentao He, Yinpei Dai et al.

PhysiCalWorld calibrates action-conditioned latent world models to instance-specific physics using in-context prompting. It encodes an exploratory interaction clip from the test scene into a compact physics descriptor token and conditions future rollouts on it, trained end-to-end with only a latent prediction loss (no physics labels, no fine-tuning). At inference, the descriptor is extracted once and reused across rollouts.

Latent DynamicsAI summary
2026-10-22 · arXiv:2610.30129

Rethinking Video Autoencoders for Interactive World Models

Adrien Ramanana Rahary, Yutong Bai, Nicolas Dufour et al.

The authors study how video-world-model compression affects interactive latency/throughput and rollout/action quality. They introduce CADRE, an open family of causal diffusion representation autoencoders spanning multiple spatio-temporal compressions. Training the same causal diffusion world model on each compression, CADRE “forms most of the latency-quality frontier,” especially with a moving camera, and yields findings about where/how much to compress and diffusion vs perceptual regression decoding.

2026-10-22 · arXiv:2610.30136

Agentic World Action Models

Zhuoyang Liu, Dantong Niu, Botong Zhang et al.

AW-0 is an agentic world action model that integrates code-based planning with joint video and action prediction. The authors jointly generate code (skill programs/logic) and future robot behavior as three coupled streams in a mixture-of-transformers architecture. An automated curation pipeline labels demonstrations with aligned programs and consolidates routines into reusable skill libraries. Evaluations on RoboTwin 2.0 and physical bimanual tasks show higher task performance than VLA and WAM baselines.

DrivingAI summary
2026-10-21 · arXiv:2601.07823

Robotic Video World Models: A Survey of Applications, Research Challenges, Future Directions

Zhiting Mei, Tenny Yin, Ola Shorinwa et al.

This survey reviews video world models for robotics, focusing on applications (robot data generation and imitation learning, dynamics/rewards modeling in RL, policy evaluation, and visual planning), key video-model architectures, and trustworthiness challenges. The authors highlight that video models can improve photorealistic, physics-consistent prediction, but still hallucinate (e.g., physics violations) and can struggle with instruction following, while facing data and compute overhead.

Video World ModelsCode availableAI summary
2026-10-21 · arXiv:2610.26903

Beckmann World Models: Direct Terminal-Map Learning for One-Step Action-Conditioned Prediction

Adam Lee, Shobhit Agarwal, June Pyo Suh

Beckmann World Models (BWM) adapt the terminal-map conservation idea to action-conditioned one-step video prediction. They learn a direct map from noise, visual history, and actions to future frames without a pretrained generative teacher or per-condition populations of generated futures. Endpoint-aligned training (source anchoring, motion/spatial supervision, and generated-history recovery) reduces errors and improves policy-score estimation under low-cost single-pass inference.

DrivingAI summary
2026-10-21 · arXiv:2610.26923

CaMeWM: A Causal Medical World Model for Generative Tumor Evolution and Treatment Planning

Huanlong Gao, Xuelei He, Fengjun Zhao et al.

CaMeWM is a causal medical world model for support-restricted post-TACE tumor evolution and treatment planning. It forms one intervention-to-decision loop: admit a queried regimen only when fold-local empirical propensity support holds, generate a distribution of lesion-local follow-up worlds via response-grounded anchor plus residual 3D diffusion, then map sampled worlds to RMST using a response-Cox utility model and evaluate the frozen policy with cross-fitted censoring-adjusted doubly robust estimation.

DrivingAI summary
2026-10-21 · arXiv:2610.26979

Latent Projection for Latent Planning

Yoo Sung Jang, Michael Ryoo

The authors propose SIGReg-based latent projection for latent-space planning. Because SIGReg-regularized latents are noisy in directions carrying no signal, they fit a constrained normal state-dynamics matrix A and use the induced covariance to project the usual ℓ2 terminal-latent cost onto the k least-uncertain directions. This post hoc change improves planning success, especially on harder 2D navigation settings, with code released.

Latent DynamicsCode availableAI summary
2026-10-21 · arXiv:2610.27045

Physical AI: Concepts, Techniques, and Challenges

Weinan Zhang, Chunhua Shen, Tianfan Xue et al.

The authors define Physical AI as an artificial system’s capacity to perceive physical states, maintain task-relevant representations of physical relationships and action consequences, and select/execute interventions via physical interfaces using feedback under physical, operational, and computational constraints. They frame Physical AI as a persistent closed loop (perception→modeling→prediction→decision/planning/control→execution→feedback) and contrast it with digital AI, embodied AI, CPS, and digital twins.

RoboticsAI summary
2026-10-21 · arXiv:2610.27280

VPFlow: Value-Predictive Planning with Imperfect World Models for Offline Goal Reaching

Yunzhe Hu, Dong Xu

VPFlow is an offline goal-reaching method that trains latent world models using closed-form hindsight Monte Carlo (MC) value regression rather than next-state prediction. At test time it uses a value-predictive planner with a one-step flow policy and evolutionary noise-space search guided by generalized advantage (GAE). On OGBench, the authors report higher success rates than model-free baselines and latent planners.

Latent DynamicsAI summary
2026-10-21 · arXiv:2610.27364

Predictive Closure: A Set-Level Characterization of JEPA's Feature Selection

Mosh Levy, Aly Lidayan, Yoav Goldberg

The authors propose “predictive closure,” a set-level characterization of how JEPA selects which features to encode. Instead of ranking features by individual predictability, predictive closure favors feasible feature sets that maximize predictive information (mutual information between current and next values). In controlled CIFAR-10 collage experiments with small latents, the authors report that unpredictable features can still be represented if they predict other represented features, and that steering can be indirect via predictive partners.

JEPAAI summary