Researchers propose OfficeTown, a professional organization simulation framework initialized from a natural-language description of an organization; members coordinate work, delegate tasks to assistant agents, and review deliverables, while artifacts, decisions, and unresolved matters continuously shape subsequent work.
MIST separates "executing the change" from "remembering the change" in video world model evaluation, locating the bottleneck in the model's ability to set up a scene but failure to carry out the instructed change. Based on this, the proposed M3-Agent (Make the Missing Move) recursively revises the details of the input instruction, enabling a frozen model to successfully execute the instruction, with the reveal score more than tripling relative to the baseline.
The paper proposes defining the usefulness of symbols with the solving algorithm at the center, introduces the concept of "solver-sufficient symbolization," and develops a counterexample-guided symbol invention algorithm for finite-horizon piecewise-affine systems. The method can be extended to optimization objectives that use neural networks as energy functions, inducing solver-sufficient symbols from continuous optimization landscapes. Simulations and real-robot task experiments show that the resulting symbols reduce solving cost while maintaining solution quality, and improve interpretability and manipulability.
The MaSS framework proposes a structural representation pathway explicitly decoupled from contextual features, using a new scene-structure discrimination objective to learn layout-aware structural representations usable for depth decoding. Trained individually and jointly on four diverse datasets, MaSS significantly outperforms prior self-supervised methods and becomes the first method to consistently benefit from mixed-domain training; zero-shot evaluation on eight additional datasets further demonstrates its generalizability, especially in entirely out-of-distribution scenes. Source code: https://mias.group/MaSS.
The research team proposes RoboHarness, which gives LLM agents a visual-geometric control panel so they can directly understand and invoke embodied tasks. With this framework, Qwen3.8-Flash-Next surpassed the current state-of-the-art VLA models on BEHAVIOR Challenge 2025 tasks, and the implementation is open source. The authors believe coding agents have the potential to achieve generalist embodied manipulation without relying on large-scale training of vision-language-action models.
TreeOPSD proposes a critic-free, tree-aware posterior self-distillation framework for long-horizon agentic reinforcement learning. It treats each interaction turn as a tree node, branches alternative continuations from shared prefixes, and recursively backs up terminal rewards to obtain implicit turn-level credit.
The StopWatch paper diagnoses that the main reason success rates remain low after fine-tuning pretrained VLMs into navigation policies is a lack of goal grounding: the policy often fails to notice that the goal is already in view or struggles to estimate its distance. The authors propose an explicit grounding method that achieves state-of-the-art results on six benchmarks across three scene collections, and they correct annotation defects in these benchmarks. Analysis on the corrected benchmarks shows that, with ground-truth grounding, the policy nearly reaches the level of the expert that generated its demonstrations, indicating that the bottleneck is perception rather than policy learning.
NVIDIA and Simon Fraser University propose and open-source GATOR, which reconstructs complete 3D objects aligned with the scene and with poses from casual photos that require no object isolation or calibration, achieving 96.36% ADD-SB@0.1 on HANDAL and 95.63% on ScanNet++.
A Reddit user relays a robot metabolism concept proposed by a Columbia University research team, in which a type of modular robot, after its leg is damaged, discards the broken part and absorbs new parts from the ground, achieving physical self-repair, growth, and upgrading.
The first public release of Awesome Physical AI Safety, a bibliography of physical AI safety literature whose taxonomy and curation methodology were designed by Futa Waseda (NII, Japan) and whose daily curation is automated by an LLM. It covers research on safety, robustness, ethics, and trustworthy deployment for VLA, robots, autonomous systems, and embodied agents.
Professor Liu Yingxiang's team at the State Key Laboratory of Robotics and Systems at Harbin Institute of Technology (HIT) published a review on high-performance piezoelectric actuators in Extreme Manufacturing, proposing a five-dimension performance evaluation strategy covering scale, speed, force, degrees of freedom, and precision. The article explains the performance regulation mechanisms of each dimension, analyzes the performance bottlenecks of different working modes, and proposes a new approach of multi-mode fusion piezoelectric actuation to achieve high precision, large stroke, large output force, and high speed simultaneously.
A team at North Carolina State University in the US has developed a motorless, batteryless micro soft hopping robot called Ring Leaper, which under continuous infrared light can autonomously complete more than hundreds of cycles of energy storage, hopping, and self-resetting.
The paper proposes a unified causal training framework, OneWM, in which actions serve simultaneously as conditions for world prediction and targets for policy learning within a shared video-action architecture. The authors evaluate policy learning and world modeling on LIBERO, LIBERO-Plus, and RoboCasa365, supplemented by real-world policy experiments: learning the visual consequences of specified actions improves policy performance and cross-scene generalization, while action learning under the same architecture improves the fidelity of multi-view world prediction. Positive results were obtained for mutual gains in both directions.
Xidian University and Sun Yat-sen University propose AdaAnchor4D, an adaptive anchor deformation framework for 4D reconstruction of dynamic scenes from monocular UAV footage. It reaches SOTA on PSNR, SSIM, and other metrics across the UAV-Arc4D, VisDrone, and UAVDT datasets while maintaining real-time rendering. Built on 3D Gaussian Splatting, the framework uses three modules—ACFA anchor-conditioned feature aggregation, DLGD decoupled local geometric deformation, and DACW density-adaptive coordinate warping—to eliminate ghosting artifacts and preserve dynamic details.
Dexmate and CMU introduced the MeMoREval benchmark, which uses a two-stage protocol of experiencing first and asking questions later to evaluate what embodied agents should store as memory. It includes 53 tasks and 4 simulated indoor environments, and runs on the Vega 1 Pro dual-arm mobile manipulator. Results show that storing raw camera frames scored 70% and had lower write cost, surpassing text descriptions at 56 to 60%; adding scene graphs achieved the highest overall success rate among all tested models. The protocol also completed two memory retrieval demonstrations on a real Vega 1 Pro using a SLAM map and head camera.
Why it matters: The benchmark was proposed jointly by the platform developer and CMU; its frame-versus-scene-graph comparison can help with memory design choices, and the real-robot transfer offers a feasibility reference.
MonoRace, an autonomous racing drone system from TU Delft's MAVLab, flew at 100 km/h at the 2025 Abu Dhabi Autonomous Drone Racing competition, beating all other autonomous teams and three former human world champions; the drones and track were provided by the organizers.
Galbot's PASSAGE system enables humanoid robots to adapt full-body motion to cluttered spaces by crouching, stepping, or turning sideways. On the Unitree G1, it completed 50 real-world tests across 45 obstacle layouts without falling or requiring safety intervention. However, in simulation, when zero contact was required, the success rate dropped from 98.67% to 70.3%.
Xin Ye, a doctoral student at the University of Michigan School of Information, and Professor Lionel Robert published a study in the Proceedings of the Human Factors and Ergonomics Society Annual Meeting. In an experiment with 100 participants involving a parking lot security robot, they found that interactional fairness and distributive fairness significantly increased people's trust in the authority of security robots.
ETH Zurich PhD graduate Karen Antorveza Paez combined microbially induced calcite precipitation (MICP) with digital manufacturing, proposing two methods for manufacturing building components.
nematode v0.7.0 (Phase 8) is released. Experiments show that the nematode connectome's learning advantage comes mainly from gap-junction placement rather than connection strength, and that this advantage still holds in a 12-segment kinematic body. The worm learns to weathervane toward a thermal set point but does not learn klinokinesis.
The UCLA RoMeLa lab has published a PhoneBot paper on arXiv: a bipedal robot 483 mm tall, weighing 1.8 kg, with 13 degrees of freedom and about $400 in material costs. A phone mounted on the torso serves as sensor, onboard computer, and interaction interface, handling motion control, image recognition, and voice interaction; speech-to-text and large-model inference still rely on an external laptop, and training uses simulation plus left-right leg mirroring. With a used phone, the total cost is under $500. The team describes it as a lab test device with no arms that does not perform complex tasks such as carrying loads or climbing stairs.
This article systematically reviews line-structured light plane calibration methods. It notes that traditional methods (the wire-drawing method, toothed-target method, and cross-ratio invariance method) rely on physical measurements and geometric constraints, yielding a limited number of calibration points and limited precision. Methods combined with camera vision algorithms (the light-stripe center point method, vanishing point method, vector cross product method, and direct calibration method via centerline extraction) use the camera's intrinsic and extrinsic parameters to convert pixel coordinates into 3D coordinates, obtaining more high-precision calibration points and generally raising average measurement precision to the sub-millimeter level.
An electrohydraulically actuated soft robotic fish developed by the Shenyang Institute of Automation (SIA) of the Chinese Academy of Sciences achieved a measured swimming speed of 2.13 body lengths per second (about 26 cm/s) and a minimum cost of transport of 9.4 J/(kg·m), setting a speed record for soft robotic fish propelled by body and caudal fin (BCF) locomotion.
KAIST, together with Caltech and Nanyang Technological University, developed BeetleBot, a beetle-inspired soft robot that users can remotely control to walk, turn, and grasp by bending and extending a finger. The robot is driven by 3D-printed liquid crystal elastomer artificial muscles, and an onboard camera and capacitive pressure sensors relay visual and tactile information, enabling bidirectional feedback. The research was published in Science Advances; the artificial muscles' speed, energy efficiency, and power supply still need improvement, and performance still needs to be tested in real work environments.
A team at North Carolina State University in the United States has developed a mechanically intelligent soft robot that carries no microcomputer and relies on no real-time algorithms. Using only two air tubes to control inflation and deflation, it can switch among three swimming modes—jellyfish, fish, and rotor—and perform actions such as ascending, descending, hovering, and turning. Its control logic is handled entirely by its materials and structure. The results were published in Science Advances.
The NeurIPS'26 Spotlight paper MoSE3 has been open-sourced; it can predict, in a feed-forward manner from in-the-wild monocular RGB video, dense SE(3) motion in the world coordinate frame for each pixel (rotation, translation, and rigid-body grouping).
MEgoVista reconstructs hand motion when an object occludes the hand: when the demonstrator naturally grips a cup and some fingers are blocked by the cup, the system can still extend the reconstructed hand into the occluded region. The method optimizes hand motion over the entire video, combining reliable visual observations with temporal continuity and anatomical constraints, thereby using information beyond a single occluded frame. For small teams collecting robot demonstration data, this means demonstrators do not need to change their grip to expose their fingers.
AIRclub-UdeSA released a ROS 2 bag recorded on a real robot for jar_workshops Week 10: a 2-minute capture in an arena filled with boxes, containing lidar, camera, odometry, IMU, and TF data. wget download and extraction commands are provided, and the empty-arena map comes with the Week 10 mapa_laberinto package.
Researchers at Chemnitz University of Technology tested pedestrian reactions to an autonomous, manual, and teleoperated campus shuttle in three modes and at different approach speeds. Preliminary results show little difference in experience between modes, while faster approach speeds slightly reduce perceived safety and comfort. A separate experiment found that braking curves and stopping distance significantly affect how pedestrians perceive the interaction. The findings are preliminary and need to cover more vehicles and traffic scenarios.
The Embodied Data Atlas project has released v0.5.0-rc.1, adding six dedicated primary research task categories and expanding its reviewed catalog from 40 to 171 published independent families (177 records in total).
Researchers have developed a tiny magnetically controlled soft robot that can reach specific locations in the body under the guidance of an external magnetic field, leaving behind a small adhesive patch for precise drug delivery; the remaining robot can continue moving to the next target and repeat the placement. The technology could make targeted drug delivery more precise, but it is still at the research stage.
Researchers have developed a robot that can shoot artificial spider silk; the polymer solution forms hair-thin fibers in the air that can adhere to Teflon and bridge sharp broken glass. A single silk web supports 100 g, and a soft silicone gripper made from these fibers can pick up delicate flowers without damage.
Any2Track addresses the difficulty of a single reinforcement learning policy handling vastly different motions such as dancing, fighting, and getting up from a fall. It uses a two-stage approach: first training specialized teachers by cluster, then using DAgger to aggregate the teachers' behaviors into a single student network, and it presents G1 simulation comparison results.
Zhejiang University and other institutions propose Pumpire, a unified benchmark that directly evaluates the ability of image-level and video-level 3D foundation models to estimate point-to-point metric distances in reconstructed geometry, noting that previous approaches of separately evaluating depth and camera intrinsics cannot reflect this ability.
The GitHub project bhl-robustness-ladder has released the research-campaign-evidence-20261011 pre-release version, providing checksum-verified and validated source code and raw simulation evidence for preregistered follow-up research campaigns on perception, sensor fusion, and terrain. The project notes that published execution receipts do not imply a positive scientific result.
Khang Nguyen and co-author Minh Nhat Vu received the CASE 2026 Best Conference Paper Award for XFlowMP, a framework that uses Schrödinger bridges to transform random noise into trajectory distributions, generating feasible motions with velocity and acceleration based on start and end configurations.
ROMA proposes a real-world object-centric multi-sensor active perception system, in which an LLM planner determines missing information, actively collects it, and provides attribute estimates once evidence is sufficient. Unlike fixed pipelines that perceive first and then act, the system uses real multi-sensor hardware, and its code has been open-sourced.
At the tactile-themed workshop at IROS 2026, Columbia University's Yunzhu Li team presented progress roughly a year and a half after the open-sourcing of FlexiTac, an open-source flexible piezoresistive tactile sensor. The sensor uses a layered structure of FPC and force-sensitive film and can be assembled in 3 minutes after parts are prepared; the team open-sourced the complete BOM, assembly instructions, and reading software and integrated it into the LeRobot framework. Simulation adaptation requires calibrating only 3 hyperparameters of a spring-damper system.
Discrete Forcing proposes first using a single discrete forward pass to determine the overall structure of an action, then mapping it back to continuous space as the generation starting point for Flow Matching, completing fine correction through a single continuous forward pass, so that the entire action generation requires only 2 NFE.
Chelsea Finn points out that as robots become more reliable, DAgger-style data collection efficiency drops sharply; for example, a 90% success rate means a human spends more than 90% of the time as a passive observer. The Mulligan paper she cites proposes the key idea that failures often come from similar initial states, so learning can be focused on these states; in deployment, the robot performs the task and the human intervenes when needed, with each collection round focusing on the initial states where the robot fails. Across 3 real tasks and 2,550 blind tests, it outperformed uniform collection under the same budget. Paper: https://arxiv.org/abs/2610.05882 .