SLAM is beginning to use contact information between the robot and the environment as a basis for localization, supplementing pose estimation with kinematics and foot-ground constraints when vision and LiDAR are limited. KD-RIEKF from IROS 2025 was validated on the Unitree Go2, and under its test conditions, both position and attitude estimation errors decreased when the robot moved at high speed.
The visual inspection research tool LDO has released its v0.1.0 research preview on GitHub, offering an installable Python 3.11 package and the ldo-research command, integrating MegaLoc and SuperPoint/LightGlue interfaces, including causal replay and optional stereo/PnP verification; a 24-second saved-evidence demo can be reproduced on CPU.
On October 16, the 2026 AI Cloud 100 China list was released in Shanghai, and Yimu Technology was selected as a representative company in the Physical AI track. Its core product, a bionic visual-tactile sensor, uses a three-layer architecture: at the bottom, it accumulates high-quality tactile datasets; in the middle, it enables universal encoding of signals across multiple hardware devices; and at the top, it integrates touch, vision, and language. The company positions itself as the tactile perception entry point for physical AI. In July this year, it completed an E-round financing of over 1 billion yuan, with a post-investment valuation exceeding 10 billion yuan.
In a conversation with Synced, Sharpa's academic lead Zhang Kaifeng outlined three main threads for advancing dexterous manipulation: human-centered data collection, high-frequency tactile low-level control that decouples motion from interaction, and tactile-anchored cross-embodiment representation learning.
LuoWei Technology has officially launched the DM-Mini 3D ToF camera, designed for visual perception at the end of robotic dexterous hands and robotic arms. Sample applications and technical cooperation inquiries are now open. The camera body measures 60×29×30 mm and weighs about 80 g, with a depth sensing range of 0.04 m to 1 m. In the 0.5 m to 1 m working range, it can achieve 4 mm depth measurement accuracy, and it offers ultra-close-range detection starting from about 4 cm. It supports RGB images up to 1600×1200 pixels and real-time output up to 30 fps, with an optional IMU module.
On October 14, DJI released DJI Dock 4, its first uncrewed platform to use an automatic battery-swap solution. It is used with the Matrice 5TD drone and the O4 ground station. The domestic bundle is priced at 167,999 yuan and is now available for pre-order.
Shandong Zhuge Robotics completed a Pre-A funding round on October 16, jointly invested by Pudong Venture Capital, Bawei Tong Spatial Intelligence, and Shijiazhuang Quchuang Technology. Its core product, the AI LOOK blind assistant robot, relies on multi-sensor fusion and a pedestrian-pathway vision large model to perform environmental perception, path decision-making, and real-time guidance in unfamiliar public environments, without the need for pre-scanned maps or manual remote control. After the funding, the company plans to continue iterating its self-developed models and expand into scenarios assisting people with disabilities and the elderly.
At ROSCon 2026 in Toronto, Canada, Intrinsic announced that it is open-sourcing the core components of its industrial robotics platform under the Apache 2.0 license and launching Intrinsic Core, a ROS-compatible foundation software platform.
Why it matters: The article clearly lists the control, perception, and planning components included in the platform and their open-source license, so readers can judge which repetitive foundation software work it can replace.
A DeepMotion badminton robot rallied 1,452 consecutive shots with a human player, earning Guinness World Record certification. According to public reports, the robot can recognize an incoming shuttlecock in about 0.5 milliseconds, swing at 30 meters per second, and smash at over 250 kilometers per hour; its ICRA 2025 paper reports a 94.5% return success rate against a shuttle-launching machine and a 90.7% success rate in consecutive rallies with real people. The system is built on a reusable general-purpose physical motion platform, consisting of a multi-degree-of-freedom racket arm, a mobile chassis, and a self-developed robot brain.
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 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.
This article details the practical workflow for completing camera intrinsic calibration and camera-LiDAR extrinsic calibration on the HandBot-S2. Intrinsic calibration uses Kalibr with an AprilGrid calibration board, and the article provides the capture requirements and the kalibr_calibrate_cameras run command.
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++.
Cognex product manager Andrew Parrett gives an example: a classification system used to sort chicken breasts and drumsticks, when it encountered a chicken wing, did not flag it as unknown but instead forced it into one of the known categories. This reflects a real problem in industrial inspection and robotic sorting—models still produce predictions for inputs outside their training range. The article describes how manufacturers handle such cases through rejection, reflow, or manual review, and how they train and validate against changes in lighting, materials, and incoming materials.
Recently, Horn Robotics, a wholly owned subsidiary of Horn Auto Electronics, received a product designation letter from a leading customer in the robotics industry, securing exclusive supplier designation for two projects: a 3M fisheye camera perception system and an ultrasonic radar perception system, with estimated revenue exceeding 30 million yuan.
The LEGO part recognition project BrickSort has released its first public version, v0.1.0, along with a trained model, lego_model.pt, that can recognize the type, color, category, and whether a part is printed for about 600 LEGO parts. The code is under the MIT license, and the model file uses the same license because its training data (the B200 LEGO Detection Dataset) is CC BY-SA 4.0. The author notes that the accuracy printed by training is optimistic, and an independent test on photos of unseen parts has not yet been conducted.
On October 14, Poxiao Intelligence and Tinno Mobile launched SeePhys Hub, a full-body, full-modality data collection system. It consists of a four-camera head ring, two wristbands, three waist-and-leg motion trackers, and a tactile glove, and can synchronously collect first-person vision, hand pose and high-density tactile data, and full-body motion information.
On October 14, Hesai launched the Kosmo One spatial camera and the Kosmo World 3D content platform, covering spatial capture, 3D reconstruction, scene editing, and robot training workflows.
Hesai held its Kosmo global launch event in Shanghai on October 14, unveiling Kosmo, a new three-dimensional world data engine, and disclosing that its robot motion modules have begun commercial delivery, with cumulative deliveries exceeding 10,000 units. The next-generation motion module offers roughly 3x the torque density and power density of benchmark products, is 37% smaller in volume, and achieves transmission efficiency above 95%; the company expects motion module shipments to reach the hundred-thousand level by 2027, with a revenue guidance of 200 million to 300 million yuan for its strategic growth business in 2026.
foundationpose-ros2-tracker releases v1.0, its first public release, accompanied by the demo data package front_4obj_move_db3.tar.zst.
Shenzhen Leishen Intelligent System Co., Ltd. invites industry professionals to attend the 2026 4th China Marine Equipment Expo. Founded in 2015, the company is the world's only lidar company that simultaneously masters four measurement principles: TOF time-of-flight, phase method, triangulation, and frequency-modulated continuous wave. It has mass-production capability for 1550nm and 905nm automotive-grade lidar, and its products are applied across ten major industrial ecosystems including autonomous driving, robotics, and smart logistics. The company also received the 2020 National Science and Technology Progress Award, Second Class.
Mair Microvision offers an integrated "3D vision + AI algorithm" solution for intelligent loading and unloading, covering depalletizing, palletizing, and loading/unloading. It has been deployed at leading customers such as Lingzhi Technology and Lanjian Intelligent. Its soft-bag depalletizing solution uses deep fusion of RGB and point cloud and supports zero-shot recognition, with recognition accuracy reaching 99.9%; it has been applied in batch to Lingzhi Technology's depalletizing robots and exported to Ethiopia. The M4 3D camera, based on iToF technology, outputs high-resolution depth data and is used in Lanjian Intelligent's loading/unloading robots for continuous corrugated box loading and unloading.
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.
SLAM's core challenge is shifting from precise localization to long-term "memory": robots need to continuously update their environmental memory and distinguish stable structures from dynamic objects. GaME and Flow4DGS-SLAM from CVPR 2026, and DyGS-SLAM from ICCV 2025, address scene changes from the perspectives of Gaussian representation, optical-flow assistance, and dynamic reconstruction, respectively.
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.
At China International Industry Fair (CIIF) 2026, ABB Robotics launched the compact IRB 1700 and IRB 1700ID robots, covering five models with payloads from 6 to 12 kg. The company claims cycle times up to 20% shorter than comparable robots on the market, path repeatability as low as 0.03 mm, joint speeds up to 700 degrees per second, power consumption reduced by up to 15%, oil change intervals of up to 20,000 hours, and a footprint 50% smaller than the previous generation. The IRB 1700ID routes cables and hoses inside the arm body.
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.
Atom Robots is presenting its MINI parallel robot at PM SHENZHEN 2026, the South China International Powder Metallurgy Exhibition (booth 16-A96), with a live demonstration of a flexible sorting solution for tiny precision parts. The compact model fits narrow workstations and features a native vision plus integrated drive-and-control system, enabling automatic recognition, orientation detection, and flexible sorting and tray placement of scattered tiny parts. Graphical, simple programming speeds up changeover and debugging, suiting high-mix, low-volume production. The exhibition runs until October 16.
The author built the TAKTO ONE tendon-driven hand exoskeleton for a master's thesis, with one magnetic encoder for each of 12 finger joints and two IMUs driving a digital twin in a headset, without relying on camera tracking, so occluded fingers can still be measured. The CAD, PCB, firmware, and build guide are all open source, and a browser demo that requires no hardware is available at https://takto.one and https://github.com/molanocortes/takto-one.
lidarslam_ros2 has released v0.10.0, the first release since v0.9.0 of the rosbag2-to-Autoware-compatible map workflow that supports Humble and Jazzy. It adds camera coloring, a MID-360 handheld preset, and a degenerate-scenario LiDAR-inertial odometry (tunnel) research branch, and fixes the default workflow after the sota-v6 merge; release-gate benchmarks show most profiles passing, with v1.0 readiness at 8/10.
Neura Robotics has released Neuron AI, a multimodal embodied foundation model that combines 3D vision, touch, audio, robot state, and language with memory, using a Perception-Language-Action (PLA) architecture.
NEURA Robotics has released Neuron AI, a multimodal foundation model for Physical AI that combines 3D vision, touch, audio, robot state, and language with memory in a perception-language-action (PLA) architecture.
A developer receiving data on the RDK X5 via MIPI CSI-2 (DT=0x1E, pack_mode=1, 1080×2912, wstride 2160) found that the image area parses correctly as YUYV but shows wrong colors as UYVY, and that the 16-bit depth and metadata at the frame tail must be byte-swapped every 2 bytes before they can be parsed; GMSL Pixel Mode and Tunnel Mode give the same result.
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.
The author integrated real Ouster OS0 lidar data and a fixed iPhone camera into a self-built "robot dojo" digital twin, correcting the camera/LiDAR calibration over three iterations; the first approach was off by about 0.5 meters and 0.4 meters too high.
RealSense explains D585 Pro on-device human detection: the detection model runs on its Gen5 Vision SoC and outputs JSON metadata at 30 fps with class, confidence, pixel box, and millimeter-level distance, without needing a host GPU.
Why it matters: The article breaks down why on-device detection does not use neural-network ranging and provides a detection distance table for different postures, which can be used directly for safety zone design.
On October 11, the Perseverance rover completed a drive that brought its cumulative distance traveled on Mars to 28.10 miles (45.22 kilometers), breaking the 45.16-kilometer record held by Opportunity.
OSARO's AI perception software SightWorks uses a deep-learning-first approach, enabling robotic arms to recognize objects, determine grasp points, and complete complex pick-and-place tasks without per-SKU entry or manual programming.
qi-neo released active-suspension-lidar-preview v2.0.0, adding the suspension_preview_v2 package, which refactors the LiDAR road preview into interfaces and coexists with v1.