Digest

Reset

1167 papers

2026-10-30 · arXiv:2610.40401

Continuous-Action World Models: Learning Cross-Embodiment Spatiotemporal Dynamics Priors

Liangyu Li, Qingwen Liu

CoAWM (Continuous-Action World Models) learns a shared latent function space for robot actions across embodiments and execution frequencies. It uses two stages: (A) pretraining an action-function encoder/decoder with action reconstruction and consistency across temporal sampling views, and (B) aligning the learned function space to a world-model backbone conditioned on observations, language, and robot state, then decoding actions via time queries at desired execution frequencies.

DrivingAI summary
2026-10-30 · arXiv:2610.40495

LucidLyra: Feed-Forward Generative 3D Scene Reconstruction and Joint Understanding via Multi-Task Distillation

Oliver Hahn, Rodrigo Marcuzzi, Sangeetha Grama Srinivasan et al.

LucidLyra distills frozen 2D foundation-model predictions from a camera-controlled video diffusion model into a feed-forward decoder that outputs queryable 3D Gaussian Splatting scenes. From a single image, it predicts geometry, appearance, open-vocabulary semantic features, and class-agnostic instance features without per-scene optimization, manual 3D annotations, or captured multi-view data; it also extends to monocular video dynamics.

2026-10-30 · arXiv:2610.40503

Algebraic World Model: From Exploration to Proofs

Guancheng Pan, Qixin Tan, Yang Yuan et al.

AWM is a framework for learning explicit mathematical world-models of unfamiliar environments via interactive Python hypotheses, structuring them with category-theory-guided transformations, then formalizing in Lean 4 to construct kernel-checked reachability and impossibility proofs. On ARC-AGI-3, AWM’s Lean models replay certified winning trajectories on all 183 levels and generalize to unseen transitions. It also proves impossibility on two modified levels with closed Lean proofs excluding every legal winning sequence.

Latent DynamicsAI summary
2026-10-30 · arXiv:2610.40532

BURGER: Benchmarking and Understanding Robotic Generalization via Embodied Drawing

Xiangyu Xu, Bangshuai Peng, Ziyi Liu et al.

BURGER introduces a language-conditioned robotic drawing benchmark and prompt-to-trajectory pipeline that generates continuous end-effector drawing trajectories in simulation and the real world. It uses 28,200 simulated demonstrations across 282 drawing tasks to evaluate five VLA models on data scaling, language grounding, spatial relations, and multimodal behavior. The authors also report real-robot deployment on a Franka Panda arm with reactive human intervention.

RoboticsAI summary
2026-10-30 · arXiv:2610.40565

Gym-Anything for Robotics: Scaling physical simulations with digital agents

Pranjal Aggarwal, Sean Welleck

The authors introduce Gym-Anything for Robotics, a framework where coding agents create robotics simulations and an adversarial creation-audit loop checks simulations and tasks against defined criteria. They build RoboWorld: 25 simulations with 25 test and 175 training RL environments. Frontier models fail all RoboWorld test tasks (average reward 0.06). The authors propose new learning methods using privileged simulator state and release the pipeline, environments, and benchmark.

RoboticsAI summary
2026-10-30 · arXiv:2610.40607

Fight World Model: Customizable Combat through Visual Interaction Proxies

Jiahua Li, Zixu John, Tom Zhong et al.

Fight World Model (FWM) separates rule-governed combat execution from video rendering using an executable engine and a deterministic compiler that outputs a spatially and temporally aligned visual interaction proxy. The proxy separately encodes attempted actions, brief hit reactions, and persistent health with identity binding, enabling rule edits (e.g., damage scaling, immunity, lifesteal, multi-stage skills) and appearance customization without retraining the renderer.

2026-10-30 · arXiv:2610.40614

L3G-World: Real-Time Infinite World Generation with Latent 3D Gaussian Memory

Jinhyeok Choi, JoungBin Lee, Jongmin Lee et al.

The authors introduce Latent 3D Gaussian Memory (L3G), a compact 3D memory that grows with scene extent rather than generated stream length. L3G-World alternates a memory updater (L3G-Mem) and a memory-conditioned video diffusion transformer (L3G-Gen) to generate spatially consistent worlds in real time. They report 9.02 FPS on one H100 GPU and large efficiency gains versus prior explicit 3D-memory methods.

2026-10-30 · arXiv:2610.40637

world2vec: Mechanistic world models through latent dynamics learning

Tobias Schmidt, Rodrigo Gonzalez Laiz, Kriszti\'an Szuppinger et al.

World2vec learns mechanistic world models from observations and actions without using reward information. It learns latent dynamics that reproduce the true system over long prediction horizons, and the authors report “at least one order of magnitude fewer parameters” than DreamerV3, TD-MPC2, and LeWM. The method achieves state of the art control on continuous environments from pixels and accurate system identification across a benchmark of 71 attractors.

Latent DynamicsAI summary
2026-10-28 · arXiv:2610.35877

Calibrated for Which Policy? Conformal Uncertainty for Fixed World Models Under Intervention

Ayush Mittal, Dhruv Gupta

The authors study how changing the deployment policy can break a fixed world model’s uncertainty calibration when only the query distribution over pre-outcomes changes but the predictor stays fixed. Under shared conditional mechanism P(Y|X)=Q(Y|X), they identify an intervention likelihood-ratio w(X)=dQX/dPX and use weighted conformal calibration plus learned ratio estimation to restore target-law coverage; a coverage penalty depends on ratio error (via TV and classifier log-loss regret).

DrivingAI summary
2026-10-28 · arXiv:2610.36011

E0.5: From Reconstructive to Future-Predictive 3D Foundations for Robotic World-Action Modeling

Era AI Lab Team, Jianhao Lai, Zhihao Zhan et al.

E0.5 is a geometry-native world-action model built directly on a 3D foundation model (VGGT-Omega). It uses 3D geometry as the native representation for world modeling and action prediction, and explicitly predicts future 3D states (depth/point maps/camera parameters) together with action. On RoboTwin 2.0 it reports 89.90% (Clean) and 50.09% (Randomized) success without embodied-data pretraining; with pretraining on 160k+ trajectories it reports mean real-world manipulation gains.

RoboticsAI summary
2026-10-28 · arXiv:2610.36068

Rollback: Experience Backtracking in World Models for Policy Improvement

Jiayi Luo, Yang Zhang, Liliang Chen et al.

Rollback turns action-conditioned world models into counterfactual engines by “experience backtracking”: it rewinds to critical decision points, branches into alternative action continuations, and distills relative action-quality supervision from contrasted outcomes—without an additional learned value model. A distributed training loop co-evolves the world model with asynchronous real-robot execution via test-time physics grounding (LoRA-based adaptation and distillation).

RoboticsAI summary
2026-10-28 · arXiv:2610.36079

Closed-Loop Aware Planning via Imitation-Free Speed Scoring for End-to-End Autonomous Driving

Emad Aghajanzadeh, Thomas Schwarzfischer, Marc Kaufeld et al.

CLAP-AD is a plug-in planning head for end-to-end autonomous driving that learns path by imitation but learns speed by imitation-free multi-criteria scoring (collision, traffic light, stop sign) under physically grounded safety constraints. It enables explicit test-time control of speed selection without speed-conditioned training by scoring candidate speed classes and selecting a feasible one.

DrivingAI summary
2026-10-28 · arXiv:2610.36121

SuperJEPA: Pretraining with a Supervision-Augmented Joint-Embedding Predictive Architecture for High Energy Physics

Dylan Sheldon Rankin, Ho Fung Tsoi

SuperJEPA introduces a joint-embedding predictive architecture (JEPA) augmented with a supervised classification objective for high energy physics. The authors report improved downstream jet-tagging performance over supervised pretraining alone across multiple tasks, model sizes, and pretraining dataset sizes, with gains persisting for both frozen and fine-tuned encoders.

JEPAAI summary
2026-10-28 · arXiv:2610.36371

BoxWorld: Let World Engines Control and Video Generators Render

Zike Wu, Ahmed Hatem, Chintan Trivedi et al.

BoxWorld separates interaction reasoning from generative rendering for real-time interactive video worlds with user-defined behaviors. A coding agent builds a Box World Engine that updates object states via a minimal Object-Centric State Interface (layout boxes, persistent attributes, dynamic status). A Box-State Generative Renderer then autoregressively renders the evolving state stream, trained with a three-stage strategy and evaluated with PairedStateBench plus video-quality benchmarks.

2026-10-28 · arXiv:2610.36420

SO-101 Bench: Accessible Spatial Manipulation

Truman Hickok

SO-101 Bench is a simulated, Isaac Lab digital-twin benchmark for language-conditioned spatial manipulation on a low-cost SO-101 arm. It uses 138 objects with seen/unseen splits and evaluates 916 feasibility-checked episodes plus a shared 3,629-demonstration fine-tuning setup. The authors report the twin preserves sim2real policy rankings (Spearman ρ=1.00), though it overestimates success.

DrivingAI summary
2026-10-28 · arXiv:2610.36493

AC-JEPA: Anchored Cross-Modal Joint Embedding Predictive Architecture for Visual-to-Symbolic Translation

Chaofei Hong, Ziyang Chen, Baochen Wang et al.

AC-JEPA integrates cross-modal latent prediction into an SFT-trained image-to-SMILES VLM. It uses bidirectional masked-state JEPA in both visual and symbolic directions as an auxiliary objective while keeping a causal SFT “anchor” for token-level supervision. The authors report improved accuracy on WildMol Simple with clean-pair training and further gains with latent self-training using pseudo-labeled pairs.

JEPAAI summary
2026-10-28 · arXiv:2610.36624

Adapting Video Models for Robot Control via Motion-guided Asymmetric Prediction

Taeyoung Kim, Junyoung Sung, Heeseung Kwon et al.

The authors study how to adapt World Action Models (WAMs) for robot control using pretrained video generative models. They argue that standard visual supervision treats regions uniformly, despite differing relevance to control. They introduce Motion-Guided Asymmetric Prediction (MoGAP), which allocates higher diffusion noise to dynamic regions and uses static regions as context via a training-only auxiliary path. MoGAP improves robot policy success on GR-1 tabletop and helps in real-world OpenArm object relocation.

DrivingAI summary
2026-10-28 · arXiv:2610.36662

SWIM: Learning What Matters from Behaviors for JEPA World Models

Jiaming Hu, Shi Bo, Yan Zheng et al.

SWIM learns which simulator state coordinates matter for JEPA-style planning from offline behavior. It trains a gated inverse-dynamics model to predict logged actions from current states and observed future displacements, using it to define a learned weighted state metric for state-supervised training. The inverse-dynamics model is only used during training, and SWIM improves planning success over uniform state weighting across five simulated robotic manipulation tasks.

JEPAAI summary
2026-10-28 · arXiv:2610.36729

Hy-Nomas: Towards Free-form Interactive World Generation

Qiaoling Chen, Jiaxiang Cheng, Xueqing Deng et al.

Hy-Nomas is an 8B-parameter video world model with a diffusion backbone that generates variable-length video segments autoregressively across multiple interaction turns. It follows free-form text instructions and continuous camera control using structured, action-aware causal attention and a vision-language Prompt Enhancer that converts the user request and cached interaction history into conditions for the next segment.

2026-10-28 · arXiv:2610.36743

Readout and Operating-Point Sensitivity in Event-Stream Candidate Selection

Yifan Shuai

The authors evaluate offline, action-defined event-stream candidate selection on DailyDVS-200 with participant-disjoint splits and sensitivity analyses. They report that representation readout (patch mean vs CLS) and validation-threshold operating points can change model rankings: e.g., DINO-S CLS captures 101/132 with 40/388 false triggers, while patch mean captures 78/132 with 35/388. Lightweight supervised encoders are competitive and the authors find no established JEPA-specific advantage.

DrivingAI summary