How can 3D generation preserve both fine detail within each part and coherence across the complete object? STRUCTPARTS reuses holistic 3D priors to generate explicit, editable parts without losing the structure and detail learned by the original model. Its dual-frame coordinate interleaving mechanism processes each shape latent from both part-local and object-global perspectives, while a segmented mesh scaffold provides structural guidance in the model’s native 3D representation. With lightweight fine-tuning, the framework supports part-decomposed generation, part-mesh refinement, and part-level editing while reducing training requirements compared with prior approaches. Title: Seeing Both the Forest and the Trees: Reusing Holistic 3D Priors for Part-Decomposed Generation Project URL: zhao-yiqun.github.io/structp… Paper: arxiv.org/abs/2610.21850v1

What if dynamic 4D scenes could be edited as easily as they are described? Edit4D introduces a text-controlled framework that separates a scene’s static background from its dynamic character. Stage-specific prompts guide background generation, character animation, 3D/4D reconstruction, composition, and global refinement—helping preserve spatial structure and temporal motion while enabling object-level editing and controllable camera movement. Title: Edit4D: Editable 4D Scene Generation with Textual Control Project URL: chanwo0kim.github.io/Edit4D/ Paper: arxiv.org/abs/2610.21668v1

What if a 3D language map could become better at finding specific user-defined targets—without losing its ability to understand new words? This work introduces a training-free, target-aware refinement method for already-distilled 3D Gaussian language maps. A text-promptable segmenter identifies each target, then the method injects just enough target-language information into the corresponding region to make it reliably detectable, while leaving non-target regions unchanged. A centering step also prevents this injection from making cosine-similarity queries insensitive to the text prompt. The approach requires no coefficient tuning and transfers to existing field-based and object-level language maps. Across three benchmarks, it improves open-vocabulary segmentation and language-goal localization, with a real-world demonstration of a legged robot navigating to natural-language goals. Title: Training-Free Target-Aware Refinement of 3D Gaussian Language Map Project URL: lab-of-ai-and-robotics.githu… Paper: arxiv.org/abs/2610.21619v1

When a small object moves quickly, a dynamic 3D model can lose track of it—not necessarily because it lacks enough Gaussian primitives, but because those primitives are placed in the wrong locations and times. “Dusting the Phantom” introduces two allocation strategies for dynamic Gaussian representations: Seed Injection adds Gaussians to under-reconstructed regions using temporal changes, reconstruction errors, and multi-view geometry, while Temporal Biased Relocation guides moved Gaussians using time-resolved optimization signals. The approach keeps a fixed Gaussian budget and improves reconstruction in highly dynamic regions, including small fast-moving objects and fine appearance details. Title: Dusting the Phantom: Gaussian Allocation for Fading Dynamic Content Project URL: minhyeok2001.github.io/Dusti… Paper: arxiv.org/abs/2610.21302v1

What should an autonomous car do when a navigation instruction conflicts with what it sees on the road? SafeDriveVLA introduces a benchmark for evaluating navigation following and conflict awareness together, including adversarially unsafe instructions. The method separates conflict reasoning from action generation by predicting whether an instruction should be followed, questioned, or rejected. It also uses a navigation-conditioned world model to imagine what would happen if the vehicle strictly followed the instruction—helping assess whether that behavior is safe before acting. On Bench2Drive, SafeDriveVLA achieves an 88.16 Driving Score and 68.64% Success Rate, improves closed-loop language following by 49.3%, and reduces collisions under unsafe instructions from 60 to 38 compared with the best SimLingo variant. Title: SafeDriveVLA: Navigation-Conditioned World Model Dreaming for Conflict-Aware End-to-End Autonomous Driving Project URL: safedrive-vla.github.io/Safe… Paper: arxiv.org/abs/2610.20609v1

Fine-tuning a real-world mobile manipulator is difficult: its mobile base and arm must coordinate across a high-dimensional action space, while robot interaction data is costly to collect. HiRe-MoMa addresses this with a hierarchical residual policy that keeps a base policy frozen and learns targeted corrections for the mobile base and manipulator arm. Sparse human interventions—covering less than 20% of training—help guide exploration and improve value estimates. Real-world experiments on a TIAGo++ show improved task performance with around 30 minutes of online interaction, across different base-policy architectures. Title: Hierarchical Residual Policy Learning for Real-World Mobile Manipulation with Sparse Human Guidance Project URL: hire-moma.github.io/ Paper: arxiv.org/abs/2610.20607v1

Can we generate humanoid control policies directly from their neural-network weights? This work explores that question with a latent diffusion model that learns the distribution of policy weights and generates new control networks for the Unitree G1 robot. The generated networks are deployable in the real world and produce motions including walking, dancing, and spin kicks. Beyond generation, the study finds that first-layer weights consistently reflect the robot’s kinematic structure across both reinforcement learning and imitation learning—offering insight into how embodiment-related knowledge is encoded in learned policies. Title: Modeling and Understanding Neural Weights of Humanoid Policies with Diffusion Project URL: humanoid-weight-diffusion.gi… Paper: arxiv.org/abs/2610.20606v1

What if driving actions could be learned directly from how the world moves? VITAL (Vision-Induced Tokens for Action Learning) derives a visual motion scaffold from future images, then encodes the remaining trajectory error as a compact residual. This lets the action model focus on the dominant maneuver while learning precise metric corrections on a separate, normalized scale—without generating future images at inference. The same residual representation also supports bounded reinforcement post-training while keeping the visual scaffold fixed. With one action denoising step, VITAL reaches 90.8 PDMS on NAVSIM v1 navtest—within 0.13 points of its ten-step score—at 3.29× the end-to-end inference speed. It also achieves 91.1 EPDMS on NAVSIM v2 navtest and 38.6 on navhard without retraining. Title: VITAL: From Visual Motion to Precise Action Project URL: worldbench.github.io/vital Paper: arxiv.org/abs/2610.20604v1

What if imperfect robot demonstrations could teach recovery instead of being discarded? Graph-Reweighted Behavior Cloning (GR-BC) identifies the action chunks that move a long-horizon manipulation task closer to success—even within trajectories containing detours, hesitations, failed attempts, or human interventions. GR-BC builds a task-conditioned graph from successful and failed rollouts, then weights actions by how much they reduce the remaining steps to a reachable goal. This reuses recoverable behavior without reward models, critics, or progress labels. Across four precise manipulation tasks and three VLA backbones, it improves supervised fine-tuning and reaches 91% success after a few offline iterations. Title: Graph-Reweighted Behavior Cloning with Human Interventions for Precise Robotic Manipulation Project URL: gr-bc.github.io/ Paper: arxiv.org/abs/2610.20504v1

What does it take to render high-quality novel views from just a few images? InfiniSplat v2 analyzes why feed-forward 3D Gaussian Splatting struggles under large viewpoint changes and introduces architectural and training improvements to address these limitations. The method decouples geometry and Gaussian attribute prediction with a frozen geometry branch, uses ground-truth camera poses to align predicted Gaussians during training, and adds large-angle target-view supervision. Across six datasets, the same trained model achieves the best mean LPIPS and DISTS among evaluated methods in both pose-conditioned and pose-free settings; in the pose-conditioned setting, it reduces mean LPIPS by 25.8% and DISTS by 32.9% versus the strongest evaluated baselines. Title: InfiniSplat v2: Analyzing and Improving Feed-Forward 3DGS from Any Views Project URL: zju3dv.github.io/InfiniSplat… Paper: arxiv.org/abs/2610.20341v1

What if an AI world model could predict not only what each player sees, but how independent actions reshape the shared world—including events outside their current view? PAWA introduces a dataset and framework for action-conditioned panoramic world modeling in Minecraft. It pairs synchronized player POV video, world-stabilized panoramic video, and aligned actions, enabling models to learn both local perspectives and the broader shared environment. Its dual-tower architecture connects the two views through direction-aware cross-attention, while a scalable data pipeline supports more than 64 simultaneous players and flexible re-rendering of recorded sessions. Experiments show improved POV prediction quality over existing multiplayer world models alongside high-fidelity panoramic predictions. The dataset and collection pipeline are open source, providing a foundation for research on multi-agent world modeling. Title: PAWA: Panoramic Action-conditioned World Modeling for Multiple Agents Project URL: pawa.siki.moe Paper: arxiv.org/abs/2610.20240v1

What if a robot could get close-up detail without a wrist camera? FoveaVLA replaces the wrist camera with learned foveation actions. Alongside each action chunk, the model predicts which interaction region to inspect next, then receives high-resolution crops from external camera views alongside the full scene. This preserves both global context and the fine detail needed for manipulation. Built on π₀.₅ and trained with RoboInter annotations, FoveaVLA-Single reaches 70.8% success with one external RGB-D view, while FoveaVLA-Dual reaches 77.1%—compared with 52.1% for a wrist-equipped π₀.₅ baseline. The gains are especially observed with distractors, cloth manipulation, and far-away objects, while matching the baseline on plain pick-and-place. Title: FoveaVLA: Foveation Actions Replace the Wrist Camera in VLAs Project URL: tml.stanford.edu/FoveaVLA/ Paper: arxiv.org/abs/2610.20064v1

Robot policies often use the same parameters across every task and environment—even when different operations require conflicting adjustments. This can leave policies poorly adapted to the situation at hand. MorphAct introduces contextual parameter generation: as the robot’s interaction context changes, coupled generators produce low-rank updates for both vision-language and action modules. Trained directly through the native action objective, the method adapts without operation labels or expert adapter targets, while preserving the connection between visual representations and action decisions. Across three vision-language-action backbones, MorphAct improves results on LIBERO and Meta-World by up to 16.30%, with real-world gains of up to 14% over π₀.₅. Title: MorphAct: Robot Policies Morphing with the Worlds They Act In Project URL: mint-sjtu.github.io/MorphAct… Paper: arxiv.org/abs/2610.20019v1

What if an interactive world could remember what happened, evolve over time, and remain consistent across different observers? “Generative World Simulation with DAO and JING” presents a framework that connects a persistent world engine, DAO, with JING, an egocentric audiovisual experience model. DAO tracks characters, objects, spaces, relationships, actions, and world time; JING uses that evolving state—along with each observer’s viewpoint and history—to generate their ongoing experience. The approach supports worlds where users and agents can explore, manipulate objects, and communicate with characters, while their actions shape what happens next. Title: Generative World Simulation with DAO and JING Project URL: xgenlabs.ai/research/generat… Paper: arxiv.org/abs/2610.19726v1

What if one learned model could both predict how the world changes and choose what to do next? LeWAM introduces an end-to-end JEPA for control that jointly learns latent-state dynamics and action generation, without reconstructing pixels. Its Gaussian-constrained latent space supports both direct policy execution and efficient planning: the model can sample candidate actions and evaluate them using its learned dynamics. In experiments, LeWAM raises success rates on contact-rich and long-horizon tasks while planning 32.3× faster than prior JEPA-based visual planning methods, and remains competitive on behavior cloning in simulation and physical robots. Title: LeWAM: End-to-End World and Action Modeling with JEPAs Project URL: le-wam.github.io/ Paper: arxiv.org/abs/2610.12818v1

Long-tailed robot datasets can cause shared vision–language–action policies to favor common “head” tasks, even when those actions are poorly suited to under-represented “tail” tasks. Evidence-Guided Counterfactual Learning addresses this bias by identifying task-relevant visual evidence, removing it to create counterfactual observations, and using the resulting predictions to more strongly connect instructions, visual evidence, and actions. The framework adapts supervision to each task’s demonstration count, requires no additional data or region-level annotations, and adds no deployment-time computation. Experiments across simulated benchmarks, multiple VLA architectures, and real-world robots show consistent gains on tail tasks while maintaining reliable head-task performance. Title: Evidence-Guided Counterfactual Learning for Long-Tailed Robotic Policies Project URL: krlozhilderman287-creator.gi… Paper: arxiv.org/abs/2610.12861v1

What visual information do World Action Models need to manipulate effectively? A new study systematically compares visual representations, integration mechanisms, and feature combinations—separating the effects of each design choice. The findings show that strong general-purpose visual representations can provide broad robustness gains, while channel-wise joint prediction achieves the highest average success rate among the evaluated integration methods. Additional visual sources help only when they contribute task-relevant information beyond what is already available; the same source can improve one combination but hurt another. The resulting R²-WAM reaches 84.7% success on RoboCasa GR1 Tabletop and 92.8% on RoboTwin 2.0—5.9 and 2.2 percentage points above VAE-only baselines, respectively. The results point to stronger general-purpose representations and better integration mechanisms as priorities for advancing WAMs. Title: What Visual Representations Do World Action Models Need? Project URL: kailiu18.github.io/R-2WAM.gi… Paper: arxiv.org/abs/2610.13024v1

What if robot manipulation data could be generated without human teleoperation—while remaining physically grounded? Grounded-Dreamer combines video world models with physics simulation to create visually realistic, action-labeled robot trajectories. Starting from a small set of real images, it reconstructs task-relevant objects and camera geometry, generates candidate rollouts, and filters them by replaying the trajectories in simulation. The resulting synthetic data is then used to train real-world robot policies, offering a scalable path to learning from diverse manipulation experiences without human demonstrations. Title: Grounded-Dreamer: Robot Learning from World Model Synthetic Data with Physical Grounding Project URL: fffstrong.github.io/Grounded… Paper: arxiv.org/abs/2610.13164v1

What if portrait video generation could pay closer attention to the parts that matter most—while keeping speech and motion in sync? PAGE (Parts-Aware Guidance) is a backbone-agnostic framework for audio-driven co-speech gesture generation from a single image. It improves local details around regions such as the lips, eyes, and hands through Parts-Specific Reweighting, then uses Consistency Classifier Guidance and Differential Guidance during inference to detect audio-visual mismatches and refine facial and body motion with less interference. The work also introduces CAS, a large-scale Chinese benchmark with broader identity, clip, and scene diversity for evaluating this task. Experiments on PATS and CAS show improvements in perceptual quality, gesture diversity, and audio-visual alignment. Title: PAGE: Parts-Aware Guidance for Co-Speech Gesture Portrait Video Generation Project URL: aigc-explorer.github.io/PAGE… Paper: arxiv.org/abs/2610.13301v1

What if a humanoid robot could learn tennis without requiring extensive, task-specific motion-capture data? ATHLETE combines generic locomotion data with tennis motions extracted from ordinary videos. Its Trajectory-Guided Motion Matching method builds diverse whole-body movement references, which are distilled into a reactive policy that uses only the robot’s and ball’s observations at deployment. On a real humanoid robot, ATHLETE achieves 90.00% forehand and 86.67% backhand hit rates, and supports multi-shot rallies with human players. Paper and videos: anonymous.4open.science/w/AT… Title: ATHLETE: Learning Reactive Humanoid Tennis via Trajectory-Guided Motion Matching Project URL: hazintihoa.github.io/athlete… Paper: arxiv.org/abs/2610.13572v1