cycloidgen v7.10.0 has been released, adding assembly and fastener modeling to the cycloidal reducer design tool. Bearing seats are now bored to the bearing's actual outer diameter and given a lip; new additions include sleeves, washers, ISO 4762 screw and ISO 4032 nut tables, a 0.5 mm default edge_chamfer, and a 0.3 mm first-layer relief for FDM-printed parts.
BAAI, together with Inverse Matrix, Peking University, and vLLM, released PhysisServe, described as the first physical AI inference system that balances intelligence and high-frequency throughput, enabling world models with tens of billions of parameters to output in real time at kilohertz frequencies, one to two orders of magnitude faster than existing frontier solutions.
On the RDK X5 development board, the official stereo depth estimation model is only StereoNet (IGEV/GRU architecture); there is no smaller ready-made network. The available options are compiled versions such as V2.4_int8 (640x352x3x2, 23fps) and V2.4_int16 (15fps).
At booth F008 in Hall 8.1H at the 2026 China International Industry Fair, Elite Robots is demonstrating its self-developed embodied large model Yuanqi Primo coordinating composite robots, wheeled humanoid robots, and collaborative robots in collaborative operations. The on-site scenarios include optical module insertion and removal, teaching-free welding, automatic drone battery swapping, and coffee kitchen applications, among which the phone gluing and tightening solution gives metrics such as gluing accuracy <±0.1 mm, repeat positioning ±0.02 mm, and a single-station cycle time <10 s.
Newton has released v1.7.0rc2, the second release candidate for Newton 1.7, expanding solver diagnostics, camera rendering, viewer interaction, and rigid/deformable body coupling workflows. The MPM material scaling fix changes existing tuning; parameters such as mpm.young_modulus must be halved and retested. URDF damping and USD unit compatibility modes must be adjusted according to the migration notes. The simulation stack is upgraded to MuJoCo 3.15, and CUDA 12 GPU support is deprecated in this version.
The author connected their own YAM robotic arm to the Astra (GPT6) model, letting people try a robot manipulation model on real hardware, and launched a remotely controllable playground website and API. Astra uses camera images as input to control the gripper based on prompts, performing well zero-shot across a variety of manipulation tasks. The website targets curious users who want to give it a try, while the API targets advanced users who want to run their own experiments.
A developer built a remote access interface for their YAM robotic arm, connecting it to the Astra (GPT6) model so others can try manipulation models on real hardware without buying their own equipment. Astra takes camera images as input and controls the gripper based on prompts, performing well in zero-shot execution across a variety of manipulation tasks. The website is aimed at visitors who want to give it a try, while the API is for users who want to conduct in-depth experiments.
Under the theme "AI, Embodied," KUKA appeared at the 26th Shanghai Industrial Fair in 2026, launching the "KAI" intelligent platform and introducing Visual Components CN, the KUKA LBR NEXUS collaborative robot, embodied products, and new automated welding solutions.
LunCoSim has released a nightly test build dated 2026-10-14, with installers for Windows x86_64, macOS Apple Silicon, macOS Intel, and Linux x86_64. A built-in auto-updater can check for and download the latest nightly build, and users can have an AI agent read the bundled documentation, skills, and examples to build and validate missions. The official notes state that these unsigned nightly builds are intended for testing only.
tatolab's streamlib releases v0.46.0, including one breaking change: frames held by one node in the engine will no longer cause frames from other nodes to be dropped, and the shared memory pools are now bounded by device memory (#2741). The change is also fixed as a Bug Fix, with a version date of 2026-10-13.
tatolab/streamlib released v0.46.2 (2026-10-13), including one engine feature update: a native node's one-off acquisitions are now counted per shape and refused by name at 16 (#2748). This is a minor version update with no other changes recorded.
At the 26th Shanghai Industrial Fair 2026, KUKA launched the "KAI" intelligent platform and introduced Visual Components CN, the KUKA LBR NEXUS collaborative robot, embodied intelligence products, and automated welding solutions.
Changlan Technology partnered with Galaxis to upgrade its traditional manual forklift warehouse into a 'goods-to-person' model using a four-way pallet shuttle rack system plus AGV transport, with 15 rack levels and storage for no fewer than 1,300 pallets. After the project went into operation, it saved more than 2,200 square meters of warehouse space, raised assembly line efficiency by more than 80%, and achieved daily inventory clearing and settlement. Three software systems—OES, PES, and RCS—handle task decomposition, shuttle scheduling, and AMR fleet traffic control, respectively, supporting fault-tolerant operation that 'degrades efficiency but does not stop the line.'
Hai Robotics deployed 291 FlashPick robots, 78 charging stations, 5 inbound and 32 outbound workstations in Anta's Langfang regional warehouse, a space of over 6,400 square meters, replacing traditional person-to-goods picking with a tote-to-person model. After the system went live, storage height rose to 8.7 meters, vertical space utilization reached over 98%, storage density increased 3 times, and 120,000 storage locations were provided; picking accuracy is above 99%, inbound efficiency is 1,063 boxes per hour, outbound efficiency is 4,417 boxes per hour, daily capacity exceeds 210,000 items, and the system supports both B2B and B2C shipping scenarios.
Geek+'s RS Hyper RMS scheduling system brings floor travel, rack climbing, and workstation handoffs into unified scheduling, supporting coordinated operation of 5,000 robots, 50,000 square meters, and 3 million storage locations. The rack area locks only 2 columns and supports two vehicles climbing in the same column, while the bidirectional high-density road network increases travel efficiency by more than 30%; in testing, system throughput within 1,000 square meters reached 6,000 cases per hour, Max workstation processing efficiency exceeded 800 cases per hour, and the company says its AI ultra-large-scale scheduling capability leads the industry by 20%+.
Rockchip Electronics and State Labs held a strategic cooperation signing ceremony at the Shanghai Industrial Fair. The two parties will combine Rockchip's AI chips with State Labs' on-device multimodal model capabilities to jointly drive the large-scale deployment of industrial vision intelligence solutions.
ArduPilot released Rover-4.7.2 on October 11 as the stable/official version for rovers and boats. An official release announcement link is provided; specific changes are not listed.
ArduPilot released Sub-4.7.2 on October 11 as the stable/official version for submersibles (ArduSub); the official release announcement is available on the GitHub Releases page. The original text did not provide specific changes or new features for this version.
ArduPilot released Tracker 4.7.2 on October 11 as the stable official version of AntennaTracker. The official release announcement has been made public on GitHub.
ROS and Gazebo PMC members meet with community members every other Friday morning. This session is on October 23 at 9 AM Pacific Time. Participants can bring questions, bugs, or robotics projects, seek help, and show their work.
Humaan Studio 0.0.1-alpha.1 extends the workflow from robot assembly and simulation to training, policy management, and physical telemetry. New features include running versioned MuJoCo training experiments locally or over SSH, configuring RunPod compute and monitoring jobs, mirroring robot joint positions in 3D through a read-only Live Link telemetry bridge, and policy registration with compatibility checks.
ANet Drone4D V0.2 builds on the V0.1 multi-agent 4D simulation platform with an interactive sandbox, quantified target perception, nest orchestration and 7×24 zero-gap relay surveillance; a public demo site is now live. The release offers six classes of quantified perception targets, seven task types, a REST API and a Python SDK; 44 of 55 items passed in the second round of acceptance testing. Known issues include a limit of 2 concurrent sandboxes on the public site, short surveillance gaps for daytime random-walk targets, and a median frame time of 50 ms without a hardware GPU.
BAGEL released its first release, v1.7.1, on GitHub, covering all content on the main branch. The web version at https://bagel-ros2.vercel.app can be used without installation.
Semiconductor startup Ubitium announced a $14 million seed round led by Ibex Investors, bringing its total seed funding to $18 million. The company is developing a unified reconfigurable processor architecture; its first chip, the UB108, uses Samsung's 8nm process and contains 256 processing units, and has demonstrated signal processing, AI inference, and real-time control on the same array. It plans to tape out the UB410 commercial processor with 4,096 processing units and deliver evaluation units in 2027.
AutoStore announced its Fall 2026 product release, covering the FlexBinsPort mixed bin workstation, AutoCase case picking, Pio Pal robotic picking for small merchants, the CubeExplore and CubeStudio design tools, and CubeFleet maintenance management.
Developers reported on the forum that submitting a task for an S100 quantized model failed and was not submitted to the cloud. Respondents pointed out that the official workflow is to use hb_compile on a development machine or inside Docker, which directly produces a .hbm file, with no cloud submission step. They also requested error logs, the quantization entry point (web page or command line), the ONNX export script, the opset version, input dimensions, and other information to determine whether the issue is a model/configuration problem or a platform task problem.
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.
qi-neo/active-suspension-lidar-preview released v3.0.1; the code of the suspension_control_v3 package is unchanged, and its version remains 3.0.0.
AutoStore released its fall 2026 software and hardware updates, covering FlexBinsPort mixed-tote picking, AutoCase case picking, Pio Pal robotic picking via RaaS, CubeExplore ROI modeling, and the CubeFleet maintenance tool.
Yu Hao proposes that evaluating on-device large-model chips should shift from TOPS to comprehensive metrics such as effective memory bandwidth, energy efficiency, and tokens/s. A 3D-DRAM stacked cubic systolic TPU architecture measured about 80 tokens/s for INT4-quantized Decode of a 3B model at an average power consumption of 5W, with token energy efficiency of 15 token/W and bandwidth utilization of about 80%. The related 3D-TPU solution is at the sample stage, and Maitexin's AI agent MetaClaw has been adapted to Qwen, MiniCPM, and Gemma and is undergoing pilot applications.
visose/Robots has released version 2.4.4, which fixes an issue with loading continuously rotating external axes with unbounded joint limits. This version can only be installed through the package manager in Rhino 8 using the _PackageManager command; see the readme for more details.
AMD held a half-day hands-on workshop at ROSCon China 2026 for robotics developers and ROS users, covering and practicing an embodied AI development workflow that spans cloud training, simulation acceleration, on-device inference, and ROS 2 deployment.
Rerun has released version 0.39.0, a database and visualization toolbox for multimodal and temporal data. Key updates include support for text in any language and color emoji, the Rerun Viewer in marimo notebooks, Raw Bayer images, PLY mesh and 2D point support, as well as an experimental audio view (AssetAudio archetype and AudioView) and ray-stepped 3D volume rendering.
The nano-ros project releases nros 0.5.0-nros2; the nros CLI uses codegen version 14.
A forum reply addressed the problem on RDK X5 where pack_image.sh in x5-rdk-gen requires sudo, while the server only grants ordinary user permissions. The cause is that the script needs loop mounting and chroot to package the rootfs partition, and the official documentation provides no root-free approach.
DART has released version 6.20.0; a 3,000-shape scene runs 18.9x faster than 6.19.4 under the ODE detector, resting bodies can sleep and no longer sink. This version is not ABI compatible with 6.19, FCL now defaults to exact primitive-shape collision, builds require a new toolchain such as GCC 11.2 and Clang 13, and optional multithreaded solving and heap-allocation-free stepping are added.
LiteArm Studio has released version v0.23.3, which ships as a single self-contained program with no need for Python, Node.js, or a repository checkout. It supports Ubuntu/Debian (requires 22.04/12 or later, glibc 2.35), other Linux distributions, and Windows, and comes with a .deb, an executable, and a .sha256 checksum. This release fixes the issue in solo mode of distinguishing a silent Cartesian probe from an unsupported probe (#163).
LiteArm Studio v0.23.4 is released, provided as a single self-contained program that requires no Python, Node.js, or repository checkout.
nexform-tech's litegrip-cpp has released v1.0.0, a breaking change (BREAKING CHANGES): MotionEngine::set_force and LiteGrip::set_force are removed from the public API.
LiteArm Studio v0.23.5 was released on October 13. The program is a single self-contained executable that requires no Python, Node.js, or code repository, and it is offered as a .deb package for Ubuntu/Debian, binaries for other Linux systems, and a Windows installer.