Digest

Reset

1167 papers

2026-10-20 · arXiv:2610.26828

Agent World Model in the Wild: Computer-Use Agent Environment Synthesis

Zhaoyang Wang, Qianhui Wu, Haorui Wang et al.

WildAWM is a scalable pipeline for synthesizing executable personal-computer environments and cross-interface tasks for training general computer-use agents (CUAs). It supports GUI, model context protocol (MCP), and CLI together, uses only open-source models for synthesis, and applies execution-based quality gates (heterogeneous rollouts, judge consistency, and counterfactual probes) to ensure solvability, difficulty, and verifier reliability. It synthesizes 341 environment–task pairs; training on these worlds improves generalization across 8 OOD benchmarks.

RoboticsCode availableAI summary
2026-10-20 · arXiv:2610.26844

AnyGoal: On World Model Generalization Beyond Expert Rollouts

Heejeong Nam, Divyat Mahajan, Quentin Le Lidec et al.

AnyGoal evaluates world-model generalization beyond expert rollouts using goal-shifted task variants where each state factor and action stays within training support, but their joint combinations are underrepresented. The authors report broad failures across world-model families under these recombinations, with oracle-state analyses showing dynamics-level failures not explained solely by representation failure.

Latent DynamicsAI summary
2026-10-20 · arXiv:2610.26845

Train An Imaginative Agent

Jiaxin Ge, Junyi Zhang, Tianning Chai et al.

Agima (Agentic Imagination) trains an agent to imagine before acting. A single omni model proposes actions and generates imagined images of their outcomes, then reasons over imagined futures to plan. Supervised fine-tuning unifies policy and world modeling; agentic RL trains adaptive planning and branching. The authors report generalization across Maze2D, Sokoban, Pacman, and PushT, including Sokoban 2-box to 4-box transfer under sparse feedback.

DrivingAI summary
2026-10-20 · arXiv:2610.26855

Ropedia W0: Scaling Loco-Manipulation World Models across Embodiments with Human Experience

Tianqi Liu, Hanfei Zhao, Qifeng Liang et al.

Ropedia W0 is a unified egocentric loco-manipulation world foundation model spanning humans and robots. It trains on Xp-Data (∼10k hours human + ∼3k hours robot) with unified geometric/kinematic/semantic annotations, using an explicit kinematic interface: metric camera poses via Plücker-ray maps and projected skeletons for manipulation. The authors evaluate on five cross-embodiment benchmarks plus held-out human egocentric video, and report real-time minute-long rollouts after distillation for live VR interaction.

2026-10-19 · arXiv:2610.22450

DriveWorld3D Evidence Constrained 3D World Modeling and Generative Simulation for Real-World Autonomous Driving

Wei Wang (Robostreet research team Cambridge, MA, USA) et al.

DriveWorld3D is a 3D world-model framework for autonomous driving that maintains a unified world state (road surfaces, volumetric occupancy, dynamic entities, road relations) while separately representing occupancy, observability/unknown space, and evidence provenance. It uses action-conditioned future prediction and risk-sensitive planning online, and offline reconstruction plus provenance-restricted generative scenario expansion for evaluation and stress tests.

DrivingAI summary
2026-10-19 · arXiv:2610.22655

Is a Linear Probe Evidence of a Linear Representation?

Nadir Kutluozen, Christian Intern\`o, David Klindt

The authors argue that linear-probe success is too permissive evidence for “linear world models.” Asymptotically, rich feature maps—even random ones—can yield high forward linear decoding. They propose three asymmetric diagnostics (reverse predictivity, spectral concentration, and out-of-distribution generalization) and validate them in a fully observed simulation and on pretrained vision transformers using SVG-World and CelebA, finding probe success sits between random features and truly linear world models.

AI summary
2026-10-19 · arXiv:2610.22795

STRIDE: Evaluating Spatiotemporal Reasoning in Driving Edge Cases

Lieqi Liu, Rui gao, Jia-Chen Gu et al.

STRIDE is an object-centric driving benchmark for evaluating spatiotemporal reasoning in long-tail, safety-critical edge cases. It tests six task families (spatial/temporal perception & understanding, memory, extrapolation, trajectory prediction, and scene-context awareness) using linked multiple-choice and free-form visual question answering. Across 12 models, the best multiple-choice accuracy is 43.4% (GPT-6-Astra) versus ~20% random-choice baseline.

DrivingAI summary
2026-10-19 · arXiv:2610.23053

Unifying Embodied Foundation Models via a Model-to-Policy Formulation

Xinyi Hong, Mingyang Ren

The authors propose an architecture-agnostic probabilistic formulation for embodied foundation models using stochastic kernels, explicitly connecting model output distributions to action realization kernels and thus to closed-loop, task-conditioned robot policies. They show how pre-training and post-training objectives relate to deployment, and instantiate the framework for OpenVLA (autoregressive action-token prediction) and DreamZero (joint video-action generation).

Latent DynamicsAI summary
2026-10-19 · arXiv:2610.23127

BeJEPA: A Goal-Conditioned Humanoid Behavior World Model via Unsupervised RL

Liyun Yan, Haoyang Li, Yang Zhang et al.

BeJEPA is a goal-conditioned humanoid behavior world model learned with unsupervised reinforcement learning from human motion priors and robot interaction. It uses a joint-embedding predictive architecture (JEPA) with a goal-independent Dynamic Predictor for one-step latent dynamics and a goal-conditioned Successor Predictor summarizing discounted future state embeddings. The composed model guides control without long imagined rollouts, and supports online action refinement.

Latent DynamicsAI summary
2026-10-19 · arXiv:2610.23156

ORD-WM: Learning Object-Centric Rigid Dynamics for 3D World Modeling in Robotic Manipulation

Yueyang Weng, Xiaopeng Zhang, Yuxi Ou et al.

ORD-WM is a 3D world model for robotic manipulation that learns object-centric rigid dynamics. It uses an SDF-supervised vector-neuron VAE to encode each object point cloud as geometry-aware 3D vector latents, then predicts object-wise SE(3) motion increments conditioned on robot action. Instead of unconstrained latent prediction, it analytically applies rigid transformations to latent states, yielding stable geometry-preserving multi-step rollouts and efficient closed-loop MPC.

RoboticsAI summary
2026-10-19 · arXiv:2610.23181

Tree-Aware Posterior Distillation for Long-Horizon Agentic Reinforcement Learning

Heming Zou, Qi Wang, Yixiu Mao et al.

TreeOPSD is a critic-free, tree-aware posterior self-distillation method for long-horizon agentic RL. It treats each interaction turn as a node in a rollout tree, branches alternative continuations, and backs up terminal verifier rewards into implicit turn-level credit. It maps this credit to a closed-form bounded teacher distribution (relative to a frozen behavior policy) and then uses Forward-KL posterior distillation across sequential mini-batches so the gradient vanishes at the target and reverses after overshoot.

DrivingAI summary
2026-10-19 · arXiv:2610.23288

CoEvolve: Self-Evolving Code Policies from Skills to Strategies for Humanoid Robots

Zicheng Liu, Haochen Wu, Jiawei He et al.

CoEvolve represents multi-robot team coordination as executable code over fixed humanoid motor skills, and adapts it during execution using Verifier Guided Coordination Evolution. A verifier accepts/rejects each transition; rejected transitions trigger typed, verifier-scoped strategy/coordination/parameter revision from an accepted checkpoint, and verified outcomes are stored for later reuse. In simulation with two Unitree G1 robots transporting a rigid payload through corridors, CoEvolve completes 8 tasks across 15 runs (base SR@80).

DrivingCode availableAI summary
2026-10-19 · arXiv:2610.23334

LAW: Unifying Language, Action, and World Modeling for Agentic Physical Intelligence

Jiajun Cao, Kai Tang, Bohua Zhang et al.

LAW is a closed-loop VLM–WAM framework for long-horizon physical manipulation that unifies language planning, action generation, and world modeling via semantic subtask progress. A hybrid-expert VLM plans remaining subtasks, estimates/uses subtask-local values to fuse WAM action chunks, and asynchronously assesses completion to continue or advance. The WAM uses MoTE and dual-stream hierarchical memory. LAW achieves 94.21% on RoboTwin; with OpenWAM-α, 95.56%.

DrivingCode availableAI summary
2026-10-19 · arXiv:2610.23396

AWAM: Active Trajectory-Conditioned View Imagination for Robust Robot Manipulation

Jiawei Geng, Zhenyang Liu, Yuzheng Wu et al.

AWAM uses a task-conditioned virtual camera trajectory to imagine geometry-consistent observations along the planned path before acting. The predicted views are generated with a geometry-conditioned world model, compressed into a reusable predictive memory once per episode, then fused with live observations via cross-attention for diffusion-based action prediction. The authors report high success on occlusion and distribution-shift benchmarks and in real-robot trials.

RoboticsAI summary
2026-10-19 · arXiv:2610.23408

LithoWorld: An Endpoint World Model for Lithography Process-Window Recovery

Botong Zhao, Fang Yu, Xinyuan Chen et al.

LithoWorld is an endpoint world model for SEM-guided lithography process-window recovery. It uses interior representative targets to determine qualified reference appearance and correction supervision, a causal endpoint alignment branch, and a conditional flow policy that outputs a 2D correction as an action chunk. Trained on 173 FEM SEM groups, it achieves 88.5% qualified-region success and 46.0% interior success on SEM-73 test pairs.

DrivingAI summary
2026-10-19 · arXiv:2610.23423

Momentum: Persistent Dynamics for Video World Models

Nedko Savov, Stefan Maria Ailuro, Danda Pani Paudel et al.

Momentum is a two-stage video world-model pipeline for persistent dynamics beyond the target camera view. It represents an evolving scene as synchronized multi-camera videos from graph-connected reference rigs (8 cameras each). WorldUpdate generates the evolving reference-video “world state,” and WorldInteract grounds target camera video generation in that evolving memory so moving content can reappear consistently outside view.

2026-10-19 · arXiv:2610.23445

RECREG: Recovery-based Shielding for Regular Safety Properties in Reinforcement Learning

Alexander W. Goodall, Francesco Belardinelli

RecReg is a safe RL framework for enforcing non-Markovian regular safety properties using recovery-based shielding. It supports Exact, ModelBased, and ModelFree operation modes, and provides theoretical safety, convergence, and optimality guarantees (focus on tabular setting). The paper also proposes a deep RL variant using DreamerV3. Experiments using MASA-Safe-RL show RecReg achieves efficient learning and matches the theory.

Latent DynamicsAI summary
2026-10-19 · arXiv:2610.23542

Execution Before Memory: Diagnosing Memory Failures in Video World Models

Weiyang Jin, Jiwen Yu, Sihan Xu et al.

The authors argue that video world-model “memory” tests should separate (1) scene setup, (2) whether the hidden change leaks early, and (3) whether the final reveal actually shows the change. They introduce MIST to score these gates separately, showing failures often come from missing execution of the instructed change rather than forgetting. They propose M3-Agent (test-time instruction revision) and report improvements on WRBench and MBench-T.

2026-10-19 · arXiv:2610.23615

Beyond Action Prediction: Learning Search-Free Planning in Latent Space

Guanchen Du, Jinhao Du, Yanpeng Sun et al.

The authors propose HIJP, a search-free latent-space planning method that replaces test-time action-sequence optimization with a goal-conditioned inverse dynamics model. HIJP trains an inverse model and a latent predictor jointly using hindsight goal relabeling across temporal horizons plus an action-transition consistency loss, so the inverse-to-predictor composition used at test time is explicitly constrained.

Latent DynamicsAI summary
2026-10-19 · arXiv:2610.23650

JAXtari: High-Throughput and Easy-to-Modify Arcade Learning Environment

Quentin Delfosse, Raban Emunds, Paul Seitz et al.

JAXtari is an open-source, JAX-native reimplementation of Atari 2600 (ALE) that removes CPU-bound emulation bottlenecks and enables accelerator-native, high-throughput PPO training. The environment also provides object-centric observations and a flexible modification plugin system for direct semantic edits, enabling zero-shot evaluation of robustness to controlled task variations.

DrivingAI summary