The paper presents DONIC, a generalist bimanual dexterous tracker jointly trained across manipulation categories and human motion sources, and releases BiHOI, a multi-view dataset containing 1,575 sequences covering 229 objects, emphasizing sustained bimanual interaction and retargeting reconstructed interactions into robot references.
The paper proposes Memory-Centric Continual Learning (MCCL), arguing that memory should be judged by whether it improves an agent's subsequent actions rather than by fact retrieval. It includes a three-axis taxonomy (organizing memory by storage location, form, and management approach), a two-loop lifecycle, an evaluation protocol, and a memory scaling framework.
This work proposes AW-0, an agentic world action model framework that models code, future video, and robot actions as three coupled generative streams within a unified mixture-of-transformers architecture. A code expert generates parameterized skill routines and control logic, while video and action experts generate future visual states and low-level motor commands.
The study proposes ALPS, a pretrained visual state encoder that unifies three capabilities—physical state information, language alignment, and pixel reconstruction—within a single set of patch tokens. It pretrains a ViT-B/16 from scratch using only still images and text, and further refines it through cross-view self-distillation.
The Dual Exam benchmark evaluates multimodal agents and world models separately: the agent track compares 10 configurations under fixed environment response mechanisms, covering 1,088 items and 15 reporting categories, while the world model track contains 400 cases across 8 scenarios.
Echo proposes a framework that treats two task status signals, stop and failure, as two dimensions of each action chunk and denoises them together with robot actions, thereby distinguishing ongoing execution, successful completion, and failure. The method requires no frame-level progress annotations; it provides complementary supervision by constructing successful, failed, and recovery trajectories, and it lets status signals denoise faster than actions so that subsequent action denoising steps can build on cleaner status estimates. Real-world manipulation experiments show that Echo accurately predicts task status and consistently improves policy performance.
FocusVLA proposes a stage-aware multi-view routing framework that predicts future interaction intent from historical actions and uses it to dynamically coordinate global and local views during action generation. On RoboTwin 2.0, its average success rate improves by 10.10 percentage points over the strongest baseline, π_{0.5}; in real-world evaluation, the success rate rises from 71.7% to 85.0%. The framework's intent prediction and dynamic routing modules can be integrated into existing VLA frameworks without modifying the core architecture.
HiSTEVE 1.0 is a hierarchical VLA agent that achieves long-horizon autonomy and native keyboard-and-mouse control in Minecraft through a capability-aware agent harness. Its orchestration layer uses structured memory to track progress, with a VLM planner selecting subtasks within the VLA's capabilities; the control layer uses a single VLA to execute all subtasks with vision-only perception.
The study examines how closely pretrained LeWorldModel (LeWM) reproduces small action changes, measuring action sensitivity with Gram matrices at three perturbation scales across nearly a thousand states from Reacher and OGBench Cube.
The study proposes an interpretability pipeline and finds that injecting task A's activations into the action expert causes pi0.5 to execute task A's motor trajectory in 99.6% of episodes (n=1,968), and that the injected program is bound to absolute workspace coordinates while ignoring the visible scene, offering a mechanistic explanation for the perturbation brittleness reported by benchmarks.
Why it matters: The study uses intervention experiments to provide a mechanistic explanation for the brittleness of VLA policies, helping readers understand why models fail after slight scene changes.
ConfigWM proposes a configurable perception mechanism that uses task language descriptions as configuration signals to guide the mapping of visual embeddings into task-specific state spaces, selectively retaining task-relevant information and suppressing distractions. In environments with progressively increasing distraction, existing methods show a marked decline in planning success rate, while ConfigWM remains relatively stable, averaging 22.3% higher than the strongest competing method. Physical-quantity probing (Pearson r +0.365) and distracted-to-clean retrieval (+75.5%) also show consistent improvements.
Li Long, a PhD student at Griffith University, Zhao Qichao, a PhD student at Tsinghua University's Institute for Interdisciplinary Information Sciences, and teams from Fudan University and Tongji University propose Spotter: an embodied model executes continuously and autonomously while a lightweight local model supervises in parallel, and the VLM intervenes to correct errors only when they occur, avoiding the robot waiting for VLM inference at every step.
Slot-JEPA decomposes JEPA's dense patch representations into stable semantic slots and time-varying spatial masks, providing a reusable object-structured interface. Built on this interface, the Transition Memory Predictor (TMP) stores visual transitions as compact masks, retrieves context using slot statistics, and uses unbalanced optimal transport (UOT) to supervise mask prediction and evaluate candidate plans.
DragMesh: Interactive 3D Generation Made Easy, proposed by a Peking University team, has been accepted to IJCV. It targets local interaction on existing 3D assets, recovering the target part's joint type, motion axis, and joint origin through a single drag, and generating a complete, continuous joint motion trajectory without retraining a motion model for each new object.
The study proposes OverSeer, a multi-agent world model that adds auxiliary camera views fixed in space alongside agents' egocentric views, serving as a proxy for a world-centric representation to improve temporal and cross-agent consistency. The auxiliary views are redundant and nearly static, so they can be efficiently compressed with a delta-VAE that encodes only changes relative to a static reference image, reducing the number of tokens required.
NVIDIA released and open-sourced the data engine Flora, which converts raw robot video into training data such as timestamped action descriptions, process rewards, and outcome rewards; it is already used to train the GR00T series models. Flora consists of an action description model, a temporal segmentation model, and a reward model, fine-tuned from the Qwen3.5 series of vision-language models, and can label 200,000 hours of data per day on 512 GB300 GPUs.
Why it matters: NVIDIA's open-source data engine cuts robot video labeling costs to about 1/125 of a general-purpose large model and provides several benchmark comparisons, which can be used to assess the substitution space in the data labeling segment.
A robotics project uses evolutionary search and AI-generated CAD in simulation to let a robot iteratively design and test tools such as grippers to improve task performance; the robot's motions still come from a predefined task controller. The next step is to validate on real hardware using an Arctos robot and a 3D printer, and the author is seeking suggestions for automatically removing gripper fingers from the print bed.
PhysisServe, a physical AI inference system jointly released by Inverse Matrix, vLLM, the Beijing Academy of Artificial Intelligence, and Peking University, completes one chunk every 4.083 milliseconds for the 33B-parameter MiniMax H3 on 200 GPUs, about 245 chunks per second, with equivalent frame throughput approaching 1,000 frames per second and a steady-state output interval shorter than that of the 1.3B-parameter Wan2.1.
An NVIDIA research team proposes the JCA model, which can simultaneously generate video and replayable 3D motion for any given skeleton, and output a camera that registers the motion to the scene. The model keeps 3D motion in a separate model that supports arbitrary skeletons, connects to a pretrained video model through a small number of cross-attention layers, and the motion model first learns skeleton movement from motion-only data.
Tsinghua-affiliated embodied intelligence company Qianjue Technology released the predictive world model Polibrain at the 2026 Physical AI and World Models Conference. It focuses on learning physical laws and cross-embodiment reuse; rather than generating complete future frames, it predicts state changes in a task-relevant latent representation space, forming a closed loop of perception, decision-making, execution, and feedback.
Urbex is an agentic vision-language framework for city-scale 3D scenes that shifts from static ranking based on pre-generated candidate boxes to active spatial search, progressively narrowing the search area through three tools—landmark query, local zoom-in, and oblique view—and uses GRPO reinforcement learning to optimize tool-invocation strategies.
It turns out that getting a few important details right is all it takes: extracting policies using only advantage-based supervised learning can achieve state-of-the-art results in reinforcement learning (RL)!
Estimating empowerment from offline data is useful because it provides an unsupervised objective for learning the relative "value" of each state in reaching diverse outcomes. But making it computable is notoriously difficult. In our new paper, we propose a computable estimator, along with some applications!
CondenseVLA proposes adding a learnable history compression module to the decoding side of VLA policies. It uses learnable queries and multi-head cross-attention to distill the per-frame features cached by the VLM backbone into a fixed number of slot tokens, and applies auxiliary regularization that penalizes overlap among slot attention maps and decorrelates slot embeddings to reduce redundancy.
VLA-Rolling is a lightweight inference framework that reuses action tails through three stages: action-tail inheritance, per-action confidence estimation, and confidence-guided refinement based on renoising and denoising step allocation, while the underlying VLA model remains frozen.
This work presents the first fully polynomial-time approximation scheme (FPTAS) for the multi-robot motion planning problem, for a fixed number k of translating polygonal robots in a polygonal environment, covering the min-sum, min-max, and min-makespan objectives. The algorithm generalizes to polytope environments and polytope robots in any fixed dimension d, and it resolves the open problem raised by Agarwal et al. on an FPTAS for k=2 unit-square and unit-disk robots, while being simpler than existing results. The paper will be published at SODA 2027.
Thinker-WM is a unified multimodal world model that jointly models and generates RGB, metric depth, and robot actions through a shared Unified Dynamics Transformer; the same set of parameters supports joint video-action generation, world prediction, policy inference, action-conditioned generation, video-to-action generation, and inverse dynamics.
The RoboFake paper proposes using a real humanoid face robot with synchronized lip movements and expressions, combined with generated speech, to create a new type of presentation attack with no pixel editing, and tests three public presentation attack detection (PAD) models and six deepfake detectors.
TAPDreamer proposes an attack method that constructs a fixed local adversarial patch using only a public encoder, requiring no queries to the target policy and transferring across tasks and action architectures. In closed-loop evaluation, a frozen patch covering about 6.5% of the input reduced FastWAM's success rate from 97.7% to 0.0% on 40 LIBERO tasks and from 90.86% to 0.0% on 50 RoboTwin tasks, while a matched random patch still retained 81.5% and 79.2%.
FERPO (Forward Entropy-Regularized Policy Optimization) is an online maximum-entropy reinforcement learning algorithm that improves the policy by differentiating the critic's value rather than its actions, avoiding policy-update problems caused when value predictions are accurate but action derivatives are unreliable.
The EMMA framework enhances vision-language-action (VLA) policies through a generative data engine and training pipeline, addressing the high cost of collecting robot manipulation data across varied appearances. Its core component, DreamTransfer, is based on a diffusion Transformer and generates multi-view-consistent, geometrically plausible manipulation videos, supporting text-driven editing of foreground, background, and lighting while preserving 3D structure.
Gibson Virtual Environment provides a perceptual training environment for embodied agents by virtualizing real spaces, and currently contains more than 1,400 floor spaces from 572 full buildings. Its built-in "Goggles" synthesis mechanism allows trained models to be deployed directly to the real world without additional domain adaptation. The environment subjects agents to physical and spatial constraints, and the paper demonstrates sample perceptual tasks learned within it; the code, dataset, and project website are openly available.
Taskonomy proposes a fully computational method that models 26 2D, 2.5D, 3D, and semantic visual tasks in a latent space, mines first-order and higher-order transfer learning dependencies, and produces a taxonomic map for task transfer learning.
The paper proposes PhysWord, a reasoning-before-acting robotic foundation model that uses atomic action words as shared semantic anchors to align continuous representations of physical dynamics with language. PhysWord predicts upcoming atomic action words and aligns their contextual hidden states with implicit targets derived from future observations, constructing a temporally structured PhysWord space with visual and optional tactile variants.
Researchers developed a layered biomimetic robotic skin that embeds receptor-analog sensors in a compliant soft medium to mimic key structural features of human skin tissue. Using this platform, the study implemented two sensory arrangement schemes, biologically inspired and counterfactual, and quantitatively evaluated them across three passive tactile tasks: weak contact detection, two-point discrimination, and texture identification. Results showed that the arrangement reflecting key features of biologically inspired sensory organization outperformed the other configurations tested.
OmniInteract proposes a whole-body interaction control method for dual humanoid robots. Through Surface Contact-Aware Retargeting, it uses dense human-robot surface correspondences to jointly retarget both humanoids, preserving the interaction contact interfaces and their near-field geometric context.
The paper proposes a human-directed policy search workflow for multi-agent pathfinding in which coding agents iteratively propose, implement, evaluate, diagnose, and refine algorithms in code space, combining simulation feedback with explicit correctness verification.
The paper proposes InteractAgent, a multimodal large language model agent system for long-horizon human-scene interaction that completes planning, execution, observation, and reflection in a closed loop, with a single MLLM serving as both planner and evaluator and a frozen MotionLLM as the executor.
The paper proposes a dynamic camera-IMU extrinsic calibration method based on the Denavit-Hartenberg convention, enabling online estimation of time-varying extrinsics by incorporating a robot kinematic model. A companion two-step initialization strategy efficiently estimates the initial joint angles, providing a reliable starting point for subsequent extrinsic parameter optimization. Simulation experiments show that the framework can effectively accommodate changes in camera-IMU extrinsics while maintaining competitive localization accuracy.
The paper proposes the FIBER framework, which uses the executable latent space of a frozen behavioral foundation model (BFM) as a shared planning representation for heterogeneous interaction skills. Its latent policy projection (LPP) maps independently trained task policies into this representation, enabling low-frequency planning while retaining the BFM's high-frequency state feedback; a single perceptive flow planner is then trained entirely offline. The paper also proposes an adaptive-horizon BFM (AH-BFM) trained with multi-gamma forward-backward learning.