Digest

Reset

1167 papers

2026-11-02 · arXiv:2610.40798

Beyond Plausible Prediction: Safety Sufficiency in Web World Models

Zhifei Hu, Alexandra I. Cristea, Adel Bibi et al.

The authors report that plausible web world-model predictions often fail “safety sufficiency”: semantically plausible futures do not reliably preserve policy-decisive evidence for Execute/Ask/Block decisions. In a human evaluation of 3,000 predictions from ten frozen open models, 86.1% are judged plausible. They introduce SafeWebWorld and Latent Authorisation Recovery (LAR), showing evidence preservation and aggregate score recoverability can still leave fine-grained authorisation failures.

DrivingAI summary
2026-11-02 · arXiv:2610.40799

CaPtain: A Pipeline for Post-Training Code-as-Policy Robot Agents

Yijia Fan, Xilin He, Yilei Jiang et al.

CaPtain is an open post-training pipeline that turns a small open VLM into a code-as-policy (CaP) robot agent without using a VLA, frontier model, or privileged state at test time. It trains inside CaPtain-Harness via SFT on GPT-6 Astra synthetic trajectories and multi-turn RL using a simulator physical critic for turn-level credit. CaPtain-4B improves LIBERO success substantially and transfers after simulation-only adaptation.

RoboticsAI summary
2026-11-02 · arXiv:2610.40927

Agentic World Modeling for Agent Test-time Scaling without Environment Replica

Wendi Li, Huseyin A. Inan, Wei-Ning Chen et al.

The authors propose Agentic World Modeling (AWM) to enable test-time scaling (TTS) in shared, non-replicable environments. AWM reformulates next-observation prediction as multi-turn evidence acquisition using safe read-only queries to the real environment and executable computation to construct the predicted observation (without mutating state). Integrated with trajectory-level TTS, AWM enables parallel exploration and improves task success over single-shot and trained world-model baselines.

DrivingAI summary
2026-11-02 · arXiv:2610.40931

Learning Stabilizing Latent Feedback Controllers via Multi-Step Backward Losses

Paul Lutkus, Lars Lindemann, Stephen Tu

The authors study when learned latent dynamics models and encoders can support feedback controllers that provably stabilize the original system. For stabilizable LTI systems, they show that learning linear autoencoders and latent dynamics with sufficiently long multi-step “backward losses” (prediction-reconstruction) recovers structure needed for stabilization via output-feedback, provided the latent dimension captures all non-stable modes.

Latent DynamicsAI summary
2026-11-02 · arXiv:2610.40935

Let LiDAR Decide Where to Look: Geometry-Measured Token Selection for Efficient Sparse Latent World Models

Zeping Zhang, Zhuoya Zhao, Robert Lagani\`ere

The authors present Geometry-Aware Token Selection (GATS), which uses LiDAR and radar point projections to select a small fraction of image tokens before the vision backbone runs in a sparse latent world model. At 25% token budget, GATS improves nuScenes 3-sec L2 displacement error and NAVSIM EPDMS while reducing end-to-end FLOPs; at 10% budget it still beats the dense baseline. It also reports safety-critical coverage and collision-rate improvements on intent-critical scenarios.

DrivingAI summary
2026-11-02 · arXiv:2610.40947

UFMF: Radar-Guided Kinematic Priors for Real-Time 4D Occupancy Forecasting

Zeping Zhang, Zhuoya Zhao, Robert Lagani\`ere

UFMF is a real-time 4D occupancy forecaster that keeps an attention-free diagonal state-space temporal core and adds radar-guided velocity-aware dynamic priors (VADP) plus two training-only geometric constraints (world-frame dynamic supervision and class-conditioned vertical-profile supervision). It achieves 122 FPS on a single RTX 4090 and reaches 0.45 fast-displacement mIoU on nuScenes, outperforming recent occupancy world models at lower computational cost.

DrivingAI summary
2026-11-02 · arXiv:2610.41022

Density-Guided Bootstrap Boundaries for Latent World-Model Reinforcement Learning

Shashank Gupta, Pilhwa Lee

The authors propose density-guided bootstrap boundaries for latent world-model RL. Using a score-informed relative density proxy (accumulated negative divergence from a score-corrected auxiliary field), they gate the depth of imagined λ-returns with hard (stop expansion) or soft (shift toward critic) mechanisms. They preserve boundary value estimates and apply survival weights; instantiated in Dreamer V3, they prove strict contraction of the gated Bellman operator and monotonic reduction of conditional target variance.

Latent DynamicsAI summary
2026-11-02 · arXiv:2610.41030

Dreams Are Cheaper Than Flights: Curated World-Model Data Improves Aerial Detection

Rugved Katole, Vedant Patil, Jenna Kline et al.

The authors seed a 7B video world model (Cosmos) with drone wildlife frames, then use an expert to reject deformed generations and annotate accepted frames. On held-out field videos from the same drones, mAP@50–95 rises 4.3% when curated synthetic frames are added to real data and 10.3% when they replace it; on an urban benchmark it reaches +33.5%. They also estimate synthetic generation is cheaper than field labor.

2026-11-02 · arXiv:2610.41041

Before We Adapt: The Zero-Shot Generalization Bottleneck in Continual World Models

Amir Mesbah, Di Wu

The authors study zero-shot reuse of learned world models in continual RL when behavioral context is fixed but physical context changes. Using Dyna-Q and TD-MPC on controlled FourRooms tasks, they find that reuse can degrade substantially under physical-context changes, and that this depends strongly on the observation/representation. Ablations and latent-space diagnostics suggest learned representations do not automatically provide invariances needed for reuse.

Latent DynamicsAI summary
2026-11-02 · arXiv:2610.41043

Meshless Latent Dynamics Models for Model Order Reduction of time-dependent Partial Differential Equations

Lorenzo Tomada, Federico Pichi, Gianluigi Rozza

The authors propose meshless Latent Dynamics Models for model order reduction of time-dependent, parameterized PDEs. They frame mesh-agnostic decoders as Neural Fields conditioned on a latent state, analyze expressivity–stability trade-offs via Lipschitz bounds for weak/strong/pure conditioning, and show grid-independent Lipschitz bounds in the meshless case. Experiments include Burgers’, Navier–Stokes flow past a cylinder, and linear advection.

Latent DynamicsCode availableAI summary
2026-11-02 · arXiv:2610.41130

One Verdict Is Not Enough: Regression Testing Learned World Models and Their Test Harnesses

Junyu Huo, Zihao Wan, Ziqi Mao et al.

The authors report TetraState, a regression-testing protocol for learned video world models that tests paired possible/impossible items under four fixed model-native configurations. It replaces accuracy-only or single-verdict oracles with configuration-level outcome classes and outcome-transition comparisons across revisions, versioning the harness as part of the SUT. They show harness revisions can change many per-test labels while aggregate modal accuracy barely moves.

Latent DynamicsAI summary
2026-11-02 · arXiv:2610.41273

SPT-Flow: One Integration, Many Futures in Continuous-Time World Models

Mingyang Lyu, Yinqian Sun, Gongce Cao et al.

SPT-Flow is a continuous-time world model that samples one trajectory-level latent per rollout and trains via rollout matching to demonstration readout states (without intermediate ground-truth inputs). The authors argue this prevents “decorative latent conditioning” from teacher forcing. On LIBERO (30k steps), they report higher best-of-10 PSNR and lower mean LPIPS with comparable per-sample inference latency.

Latent DynamicsAI summary
2026-11-02 · arXiv:2610.41304

ART of PIV: Agentic Real-Time Particle Image Velocimetry for Active Flow Control

Antonio Terpin, Francesco Banelli, Alan Bonomi et al.

ART of PIV introduces an agentic, real-time PIV pipeline that selects a configured flow-estimation pipeline (from a library) for each image pair within a specified speed budget. Across cross-correlation, classical optical-flow, and learning-based estimators, the authors report 12.2–22.4% lower average end-point error (AEPE) versus the best single estimator at matched speed, with gains up to 1000–4000 image pair/s. They also show reduced closed-loop cost in simulation and physical flow tracking with a DreamerV3 controller.

DrivingAI summary
2026-11-02 · arXiv:2610.41330

Cross-Embodiment Motion Planning by Coupling Body-Agnostic World Model with Differentiable Ego-Dynamics

Zhidong Wang, Zirui Li, Haoran Xu et al.

Hourglass factorizes cross-embodiment local motion planning at the camera ego-motion interface: a body-agnostic recurrent world model conditioned only on ego-motion and camera parameters, coupled to each robot’s differentiable ego-dynamics (a reduced-order template) only through the implied camera ego-motion. The authors learn a generalist parameter-conditioned policy once via coupled latent imagination, and identify template parameters on new robots for deployment.

DrivingAI summary
2026-11-02 · arXiv:2610.41354

DiffWBC: Unlocking Steerable General Humanoid Motion Tracking via One-Step Diffusion Whole-Body Control

Yifei Huang, Yazhao Yang, Weijia Li et al.

DiffWBC is a 50 Hz one-step diffusion whole-body controller for humanoid motion tracking. It generates a distribution over future trajectory plans from a reference horizon in a single forward pass, then uses an inverse-dynamics head for actions. The authors report lowest local tracking error among six general trackers on held-out benchmarks and show test-time steering by optimizing generator input noise with differentiable costs, plus simulation-free failure discovery via plan dispersion.

Latent DynamicsAI summary
2026-11-02 · arXiv:2610.41654

$\nabla$Mimic: Ultra-Efficient Humanoid Control with Differentiable Simulation

Weyl Lu, Yubei Chen

∇Mimic trains humanoid motion tracking using exact simulator gradients on contact-rich control, addressing exploding/expensive gradients. It bounds rare unbounded per-sample gradients via median-relative gradient clipping and reuses simulated windows via multistep regression onto gradient-derived action targets. A GPU differentiable simulator, NablaEngine, provides fast exact adjoints for MuJoCo contact dynamics on Unitree G1. On all 40 LAFAN1 clips, it matches SuperMimic’s score 0.82 in 20 minutes (1.2 h for SuperMimic).

DrivingCode availableAI summary
2026-11-02 · arXiv:2610.41857

When the Simulator Fails, Not the Policy: Auditing Contact Artifacts in Simulated Robot Benchmarks

Alex Zhou, Junhao Wu, Xueyang Zhou et al.

Simulation robot benchmarks can miscount simulator contact artifacts as policy failures. The authors trace “contact ejection” failures to PhysX 5.3.1 contact/distance threshold behavior and audit episodes with SimuGuard. With the constant fixed (PhysX 5.4.0), episodes with artifacts drop from 12 in 100 to none, and five container-task success rises from 89% to 98%.

DrivingAI summary
2026-11-02 · arXiv:2610.41903

Statistical Foundations of Reinforcement Learning: A Non-Asymptotic Perspective

Yuxin Chen, Yuejie Chi, Jianqing Fan et al.

Monograph presenting non-asymptotic statistical foundations for reinforcement learning in tabular MDPs. The authors analyze sample complexity (PAC-style) for three sampling mechanisms—RL with a generative model, online RL, and offline RL—covering model-based and model-free paradigms, with minimax-optimal algorithm design and comparisons to minimax lower bounds. Extensions discussed include robust RL, competitive multi-agent RL, federated RL, uncertainty quantification, and RL from human feedback.

DrivingAI summary
2026-11-02 · arXiv:2610.41970

When Does Test-Time Search Help World Action Models? An Empirical Audit of Candidate Selection and Simulator Integrity

Jun Liu, Dajie Sun, Zhenglun Kong et al.

The authors audit test-time scaling for world-action models on Cosmos Policy with RoboCasa using a premise-first, pre-registered methodology. On clean data, a learned value selector is 1.66 pp worse than random; a hindsight oracle gains only 1.40 pp with an episode-level interval spanning zero. Token pruning yields at most 1.152× end-to-end speedup on the deployed path (1.253× decode-excluded), short of 1.5×, and they identify/fix a compute-dependent rendering corruption.

DrivingAI summary
2026-11-02 · arXiv:2610.42021

Nexus-VLA: Bridging Human and Robot Experiences via a Shared Hierarchical Semantic Latent Space

Biaoliang Guan, Tingyu Yuan, Wen Ye et al.

Nexus-VLA is a hybrid human-robot policy learning framework that transfers manipulation knowledge across embodiments via a hierarchical semantic latent. It end-to-end co-trains on unpaired human and robot demonstrations, progressively aligning human/robot semantics at four levels (Intent, Spatial, Motion, Interaction) and shaping a shared latent used directly for continuous action generation. The authors report 80.5% task success on RoboCasa-GR1 and strong OOD and real-world generalization.

RoboticsAI summary