Digest

Reset

1167 papers

2026-10-14 · arXiv:2610.18186

Reinforcement Learning for Large-Scale Lifelong Multi-Agent Path Finding with Shared Spatial Fields

Jean Seong Bjorn Choe

The authors propose an end-to-end average-reward RL policy for lifelong multi-agent path finding that computes a shared spatial field over the map once per timestep, then uses a goal-conditioned route operator and a destination allocation (Sinkhorn) step to decide agent movements. Trained without planner demonstrations or an online planner, it transfers to larger unseen layouts and populations, matching or exceeding best baselines across six 10,000-agent layouts while running in under 10 ms per step.

DrivingAI summary
2026-10-14 · arXiv:2610.18289

LongForcing: Scaling Temporal Manifold Matching for Long-Horizon World Rollout

Ziyu Zhu, Zhaoxu Sun, Fan Jiang et al.

LongForcing is a distillation method for long-horizon world rollout that scales teacher supervision along the temporal axis. It keeps short-horizon teacher-forcing and causal ODE-distillation, but replaces the final DMD real-score teacher with an extended-horizon bidirectional teacher that supervises longer student self-rollouts using the standard distribution-matching formulation.

Video World ModelsCode availableAI summary
2026-10-14 · arXiv:2610.18305

AffectWorld: Benchmarking Emotion-Driven Future World Prediction

Changhao Song, Yacong Chen, Hui Gao et al.

AffectWorld is a benchmark for emotion-driven future world prediction, evaluating the chain Emotion → Action → World State. It includes 1,498 everyday video samples plus a 233-sample counterfactual subset where only an agent’s emotion is intervened. The authors report evaluation of five state-of-the-art world models using eight metrics (three tasks plus counterfactual), revealing gaps in emotion-driven action consequences and emotional intervention sensitivity.

Latent DynamicsAI summary
2026-10-14 · arXiv:2610.18373

DR-VLA: Learning Human-Like Quadruped Navigation from Egocentric Web Videos via Domain-Invariant Scene Representations

Jun Tian, Yingjie Niu, Ruo Chen et al.

DR-VLA is a Mixture-of-Transformers framework for quadruped instruction-conditioned navigation learned from egocentric web videos via a domain-invariant scene representation (DISR). The authors pretrain understanding and world experts on 10,000h EgoLynx-Web without action labels, adapt to robot observations, and train a flow-matching action expert on 2,000 teleoperated clips (EgoLynx-Robot). On held-out EgoLynx-Robot, it improves QNS and reduces FHE vs competing methods, and achieves 92.1 PDMS on NAVSIM v1.

DrivingAI summary
2026-10-14 · arXiv:2610.18447

Thinker-WM: A Unified World Model with Shared Multimodal Representations for Embodied Intelligence

Thinker Team, Peng Gao, Jichao Jiao et al.

Thinker-WM is a unified multimodal world model that jointly models and generates RGB video, metric depth, and robot action using a shared Unified Dynamics Transformer with modality-specific encoders/decoders. The authors unify at the representation level so appearance, geometry, and action tokens co-evolve in a common dynamic representation under task-level conditioning. The same parameterization supports multiple world-action functions.

2026-10-14 · arXiv:2610.18568

WAVE-0: Bridging VLA Policies and WAMs with Lightweight Latent Future Prediction

Haomin Wang, Anqiong Ji, Ke Wang et al.

WAVE-0 (World-model-Augmented VLA Execution) bridges vision-language-action (VLA) policies and world action models (WAMs) with a lightweight latent world model. Using a 257M action-conditioned latent model trained directly on ground-truth robot actions with a counterfactual ranking objective, it predicts future representations that condition the action expert. It reports high success on RoboTwin 2.0 and LIBERO, plus gains on real-world tasks.

RoboticsCode availableAI summary
2026-10-14 · arXiv:2610.18570

Order-Consistent Latent Geometry for Reliable Short-Horizon Planning in World Models

Junlin Wu, Wen Fan, Dandan Zhang

The authors study latent goal planning that scores action candidates by terminal Euclidean distance in latent space. They show that preserving the correct ordering of true progress along transition-shortest rollouts is equivalent to the classical self-approaching condition. For polygonal (discrete) world-model rollouts, an acute bend can create a region where terminal Euclidean scoring prefers a transition-worse candidate. They validate the ordering failure with a controlled six-branch experiment and interpret latent dimension via an angular-capacity bound.

Latent DynamicsAI summary
2026-10-14 · arXiv:2610.18637

Grid or Nodes? Probing Object Memory in World Models Under Partial Observability

Justin McMillen, Yasin Yilmaz

The authors compare FloWM’s grid memory with a node-based Kalman tracker for object memory under partial observability in a synthetic block world. With an oracle-fed tracker, node and grid can match on FloWM’s standard motion prior (89% vs 91% out of view). But when motion is outside the prior, both degrade, and the grid’s built-in design requires many grid copies (25) to cover varied velocities.

Latent DynamicsAI summary
2026-10-14 · arXiv:2610.18645

ST-GRU PredRNN-V2: A GRU-Based Reformulation of PredRNN-V2 for Video Prediction

Augusta van Haren, Hanna Hoogen, Luca Pattavina et al.

The authors propose ST-GRU as a GRU-based replacement for PredRNN-V2’s ST-LSTM cells, using two GRU-style gates (update/reset) and merging the temporal memory state into the hidden state. They implement ST-GRU-based PredRNN-V2 (22.6M params) and evaluate on Moving MNIST using the original protocol, reporting MSE 55.1, SSIM 0.883, LPIPS 0.065, PSNR 19.53 dB averaged over ten frames.

2026-10-14 · arXiv:2610.18650

ForkCast: Predictive Recovery for Vision-Language-Action Policies via Counterfactual Futures

Chenru Wen, Huajian Zeng, Haokai Ding et al.

ForkCast is a predictive recovery method for frozen vision-language-action (VLA) policies. At each action-chunk boundary it compares counterfactual futures for three candidates—Commit (nominal continuation), Reobserve (camera viewpoint change), and Replan (planned arm correction)—using a candidate-conditioned world model and a shared task-progress head. It selects the highest-scoring candidate without a failure detector, intervention threshold, or VLA updates.

DrivingAI summary
2026-10-14 · arXiv:2610.18677

Model-Based Online Decision Making via Generative Trajectory Planning

Haldun Balim, Yilun Du, Na Li

The authors propose Generative Trajectory Planning (GTP), an online model-based framework that learns a diffusion-based generative model as a proposal distribution over action sequences. At each timestep, GTP samples candidate action sequences and refines them via beam-search-style iterative diffusion and rollouts through learned dynamics, reward, and value models, using a dynamics ensemble for trajectory-level uncertainty and optimistic exploration.

Video World ModelsCode availableAI summary
2026-10-14 · arXiv:2610.18697

UniW(A)M: A Unified Mixture-of-Experts Framework for World and World-Action Models

Jingbo Xu, Jianjie Fang, Zhuohang Li et al.

UniW(A)M is a mixture-of-experts world-(action) model built on a diffusion Transformer that unifies action-conditioned world modeling with four world-action interfaces via video–action visibility masks and CATR expert routing. It co-trains prediction and action generation, then uses PAIR to adaptively select an interface per execution chunk while keeping the backbone frozen. The authors report 88.3% average real-world success vs 76.7% for the strongest baseline.

Latent DynamicsAI summary
2026-10-14 · arXiv:2610.18699

RoboFQE: Rollout-Free Evaluation of Visual Robot Policies from Offline Trajectories

Yunfu Deng, Sangwoo Shin, Brahma S. Pavse et al.

RoboFQE is a rollout-free off-policy evaluator for visual robot policies from offline trajectories. It fits fitted Q-evaluation (FQE) without rolling out the target policy, using a pretrained DINOv3 vision backbone for high-dimensional observations and a learned terminal success detector (LRM) to provide reward signals when ground-truth labels are absent. The authors report accurate discounted return estimates and preserved policy rankings on LIBERO and CALVIN and on real robot block-stacking tasks.

DrivingAI summary
2026-10-14 · arXiv:2610.18774

GreenWAM: A World Foundation Model for Robotics

Gennady Sidorov, Valerii Kornilov, Denis Grankin et al.

GreenWAM is a robotics world foundation model that can serve both world-model and policy roles across different robot bodies. It uses a shared anatomical action space for 20 robot families and joint training in visual, action-conditioned, and world–action modes, with policies formed by splitting an action expert off the generator. The authors report top-tier WorldArena 1.0 results and strong benchmark performance on RoboTwin 2.0, LIBERO(-plus), and CALVIN, plus real-humanoid adaptation and single-step distillation.

RoboticsAI summary
2026-10-14 · arXiv:2610.18792

SEAR: Self-Evolving Agentic Robotics

Xinyi Wang, Bangzheng Li, Hanyu Wang et al.

SEAR (Self-Evolving Agentic Robotics) is a framework and benchmark for testing whether LMs improve robots through self-evolution, disentangling true learning from effects like resampling, luck, and privileged simulator state. It runs agents in three modes (code-as-policy, agent-as-policy, auto-research) and evaluates across 330 simulated and five physical tasks using seven frontier LMs, with retained “libraries” carrying skills and memories across tasks.

DrivingAI summary
2026-10-14 · arXiv:2610.18826

WorldOracle: Benchmarking Action-Conditioned World Models against Real-World Executions

Haoyi Duan, Liliang Chen, Nan Wang et al.

WorldOracle is a benchmark that tests action-conditioned world models by comparing predicted open-loop rollouts to paired real-world executions using the same initial observation and recorded low-level action sequence. The authors evaluate execution coherence along physical state correctness, interaction spatial alignment, key interaction area fidelity, and multi-view action coherence, using long multi-segment episodes with successful and failed interactions. They release datasets, evaluation code, and a leaderboard.

DrivingCode availableAI summary
2026-10-14 · arXiv:2610.18834

Game2Real-Bench: Do Toy Games Reveal and Transfer to Real-World Visual Reasoning?

Qingan Liu, Juntong Feng, Yue Zhu et al.

Game2Real-Bench studies whether toy-game visual reasoning reflects and transfers to real-world visual reasoning in VLMs. The authors build a 1,000-sample procedurally generated, programmatically verified toy-game set and align 400 real-world samples by shared reasoning requirements. Evaluating 25 VLMs, they report strong game–real model-level correlations and selective one-shot transfer effects, plus a nearly 24K-sample Game2Real-Dataset.

JEPAAI summary
2026-10-14 · arXiv:2610.18852

ForceWAM: Decoupling Future and Response Learning from Demo and Play Data for Contact-Rich Manipulation

Zijun Xu, Wendi Chen, Xinkai Wang et al.

ForceWAM is a force world action model for contact-rich manipulation that decouples learning task-compatible interaction outcomes from learning how local actions change contact. A force world model predicts future video and six-axis force/torque trajectories (visual-wrench futures) from task demonstrations. A separate inverse dynamics model uses current observations plus recorded or generated visual-wrench futures to predict robot actions, updated reactively with fresh wrench feedback. Real-robot results reach 85–95% mean success.

RoboticsAI summary
2026-10-14 · arXiv:2610.18855

Interaction Latents as Actions for Contact-Rich Manipulation

Yuhang Zheng, Yujie Zang, Yuchen Liu et al.

Contact-rich manipulation requires robots to coordinate not only where to move, but also how contact and force should evolve throughout an interaction. Yet most robot policies represent actions primarily as motion trajectories, using tactile signals as inputs or auxiliary predict…

Source abstract
2026-10-14 · arXiv:2610.18867

EmbodiedEvo: Self-improving Coding Agents for Physical Intelligence

Wenyi Li, Yizhe Chi, Youjie Zheng et al.

EmbodiedEvo is a self-improving coding agent for physical tasks that evolves an evolving harness of reusable “Skills” and “Lessons” from feedback on one Playground scene, then freezes it for Test agents on new layouts. On RoboDojo-Sim with Astra/xhigh, the authors report Average Score 44.48 and success rate 37.82%, surpassing all officially verified results. They also report cross-model transfer benefits and lower execution cost than using raw history.

RoboticsAI summary