A developer asked Reddit for help organizing the code structure of the SprayShark spraying robot. The project targets targeted pesticide spraying at home, uses a Raspberry Pi, ESP32S3, AI Hat+, and a camera to identify weeds, and the code is hosted at octotoxin/SprayShark.
A maker built a pneumatic precision press from off-the-shelf parts. The device detects the distance to the rod end (a black circular plate) with an ultrasonic sensor, and an Arduino controls the actuator's movement toward the target position. The maker used a 5/2 directional valve in an unconventional way, exhausting air from both sides of the cylinder to drive the rod, and because the effective areas at the two cylinder ports (port A and B) differ, used two pressure regulators to balance the forces and hold the rod at the target position.
Adhesive selection for humanoid robot electronic skin can follow a four-layer structure: sensing layer, substrate and body, aesthetic layer, and edge sealing, each with its corresponding adhesive type. The article presents a comparison table giving adhesive options for each layer, for engineers to consult directly and reduce trial and error in selection.
A complete walkthrough of the e-skin bonding process: starting from a piece of Ecoflex, it covers material preparation, surface treatment, and adhesive selection across three layers, then mounting the e-skin on a robotic arm and getting it running, and finally validating the bond through bending, damp-heat, and peel tests to help developers avoid half a year of detours.
The author tests MyoLeg's official pretrained policy and a spinal reflex baseline, shifting the walking control target from torque to 80 neural excitation channels, involving activation dynamics, force-length-velocity relationships, and moment arms. The article presents quantitative results on gait mechanics, impact recovery, muscle weakness, and fatigue, and summarizes lessons learned in reproducing the reflex circuit.
Centered on the autonomous tidying task of a mobile manipulator, this tutorial presents a complete derivation chain covering occupancy grid log-odds mapping, A* global path planning, model predictive trajectory tracking, and grasp-and-place control. The author reproduces the full workflow of mapping, navigation, grasping, and transport in a MuJoCo four-room scenario.
Explains the development principles of embedded Linux character device drivers and kernel modules, covering key stages from module loading to ioctl interface implementation, helping robotics developers understand the underlying mechanisms of device drivers.
EtherCAT is an industrial real-time Ethernet protocol. This article explains its three core mechanisms: distributed clocks for synchronizing clocks across multiple slaves, the process data (PDO) and service data (SDO) communication methods, and concrete master and slave implementations. For robotics developers who need high-frequency, low-latency servo control, this is a practical introduction to understanding industrial bus communication.
An in-depth analysis of the MQTT and MQTT-SN IoT protocol stacks, covering the publish/subscribe model, QoS quality levels, and the connection mechanisms for low-power devices, explaining for robotics developers how these two protocols are used in IoT communication.
This TinyWAM source code walkthrough reviews the core implementation of this lightweight World Action Model, covering data generation, visual encoding, Flow Matching, model training, and the closed-loop execution pipeline. The article explains each stage alongside the source code to help readers understand the technical principles and engineering implementation of world action models.
This practical project centers on the core Light-WAM code, analyzing how it uses a frozen video world model, an Adapter, and a State Fusion mechanism to directly decode robot actions in a more lightweight way while retaining future-prediction capability.
The Bridge-WA hands-on column walks through its source code, covering core designs such as World Teacher, Future Token, Change Map, and Motion-Flow Map, as well as WorldBridge's attention injection mechanism, and also covers training and deployment.
ST-WAM is a model for robot action generation. This article presents a hands-on analysis based on its source code, introducing the implementation of dual-space future modeling, CAIR history retrieval, and cross-branch attention, and outlining the training and inference pipelines to explain how semantic information and temporal information jointly serve robot action generation.
The Guyueju Daily Code Practice Column analyzes the WALL-WM source code, focusing on Event-Level data organization, multi-view Video DiT, layer-by-layer coupling of the Action Transformer, and the Video-Action joint mechanism, to help newcomers understand the code implementation of this world model architecture.
ROSCon China 2026 will hold the NVIDIA Physical AI Full-Stack Robotics Workshop on October 17 from 13:30 to 17:00, led on site by NVIDIA technical experts. It will cover the Isaac GR00T end-to-end architecture, Isaac Lab 3.0 perception reinforcement learning and whole-body control training, Physical AI Agent skill development, Jetson on-device model inference acceleration and memory optimization, and an industrial deployment case study from Techmach.
Mobile ALOHA is a low-cost, open-source dual-arm mobile manipulation system proposed by a Stanford research team. It adds a mobile base and a whole-body teleoperation interface to the ALOHA dual-arm teleoperation platform, synchronously collecting vision, robotic arm state, and base motion data through teleoperation, then training policies with imitation learning methods such as behavioral cloning. Among these, ACT predicts a segment of an action sequence at a time.
Before mass production, humanoid robots must pass a six-level progressive test from L0 to L5: L0 completes dynamics calibration and preliminary validation of control strategies in simulation environments such as MuJoCo and Isaac Lab; L1-L2 conduct mechanical, electrical, and control performance and NVH testing on joint modules; L3 integrates motor drives, thermal management, and communication buses; L4 conducts full-machine testing with loaded stair climbing, push-disturbance recovery, and long-cycle durability; L5 runs long-duration tests in real scenarios to compile statistics on failure modes.
After flashing a self-compiled emmc_disk.img to the RDK X5, boot fails, and fastboot reports Writing 'boot' FAILED (remote: 'cannot find partition'); the cause is that the board cannot read a valid GPT partition table.
An RDK S600 forum post investigates a tuning_bin startup error reporting a missing libevent-2.1.so.7.
Reworking OpenCV's Mat through memory-mapped files lets large images that don't fit in memory still be accessed through the Mat interface, with data still pointing to a contiguous uchar* in memory. Below the threshold, native heap allocation is still used and performance is unaffected; only genuinely large images switch paths, transparently to upper-layer callers.
Lecture 5 of Xbotics Embodied Intelligence 21 Lectures uses the SO-101 robotic arm in MuJoCo as its main thread, explaining how to break a single simulation demonstration into a simulation experiment closed loop that can be recorded, judged, and reviewed.
The Enabot 2026 Black Friday buying guide divides camera robots into three tiers: the entry-level EBO SE, the mid-range EBO SE 2, and the flagship EBO Max, and recommends choosing based on the primary use, such as watching pets, watching elderly family members, or home security, rather than looking only at specs.
A developer got Muse Gadget's cloud conversation and animated avatar interface working on the RDK X5 (Ubuntu 22.04.5 ARM64), with a 480×320 SPI small screen displaying a pixel avatar, subtitles, and touch interaction; model inference runs in the Muse cloud, and the BPU was not used for local inference.
Before bringing an AI robot pet home, first clear a safe activity area, check functions such as voice interaction and the camera, and review privacy settings; then let children observe first and start with short, simple interactions. Tell children that AI can respond but has no human thoughts or feelings and can make mistakes, and the American Academy of Pediatrics notes that AI cannot replace stable interpersonal relationships. Also set no-use periods for homework, bedtime, and meals, do not disclose private information such as names and addresses to the robot, and supervise children's interactions with real pets in the early stage.
Before buying a robot pet for your child, parents should distinguish between traditional interactive products and AI-powered ones, and pay attention to AI response quality, privacy, and age suitability. A 2026 Common Sense Media evaluation of some AI toys found that 27% of tested outputs were unsuitable for children; the result applies only to the products tested. Common Sense Media recommends that children aged 5 and under not use AI toys, and that ages 6 to 12 use them with caution.
AMR robots rely on four onboard technologies for trackless navigation: laser SLAM mapping, real-time LiDAR perception, dynamic obstacle avoidance, and real-time path replanning, with no need for floor tracks, magnetic strips, or QR codes. REEMAN delivery robots achieve laser navigation accuracy of ±5 cm, with an optional reflector code accuracy of ±3 cm; the Fly Boat model achieves QR code accuracy of ±2 cm, and a single map covers about 40,000 square meters. Its product line covers payloads from 100 kg to 1.5 tons, and since 2015 it has delivered to more than 55 countries and regions.
Compatible AMR chassis can be purchased through three channels: direct manufacturer sales, authorized distributors, or online portals. The key selection criteria are an open SDK, ROS/ROS2 compatibility, and mechanical and electrical modularity. The REEMAN chassis series includes the robot mobile chassis (50 kg), Feizhou, Big Dog, Moon Knight, and the 600 kg IronBov AGV, offering laser SLAM and obstacle avoidance plus 24/7 technical support; with complete documentation, a first autonomous run can be achieved in 2–6 weeks.
REEMAN has released a 2026 AMR chassis comparison guide covering 7 manufacturers, payloads of 50–1,000 kg, price points from $2,500 to over $45,000, and lead times of 3–14 weeks. The guide evaluates dimensions including payload class, drive architecture, ROS/ROS2 support, and SDK interfaces, and tiers the options into entry-level/R&D, mid-range OEM, integration partner, and heavy-duty categories. REEMAN's own chassis cover 50–600 kg, feature laser SLAM and an open SDK, and are positioned between research platforms and heavy industrial OEMs.
Autonomous mobile robots and drones can replace time-consuming manual inventory monitoring, improving inventory visibility and freeing staff to focus on exception handling. This report gathers automation use cases from logistics, cold chain, and retail, and covers advances in warehouse and distribution center software, as well as the trend toward more proactive inventory management enabled by digital twins and AI; it also looks ahead to a future of cycle counting that uses gamification to improve traceability, reduce shrinkage, and raise productivity.
End-effector attachments are modular work tools mounted on an AMR base that share navigation, power, and control systems with the base, turning the mobile platform into a machine that performs specific material-handling tasks. Roller conveyors, lift mechanisms, cart attachments, and custom grippers cover most warehouse scenarios, and an open SDK lets engineers integrate new attachments without replacing the entire robot, avoiding vendor lock-in. Reeman's Fly Boat robot base provides standardized mechanical interfaces and an open SDK, supporting expansion from simple transport to docking transfer and sorting pick-and-place.
At the IROS 2026 Tech Talk, Zhijian Dynamics presented its precision manipulation work for robot loading and unloading of CNC lathes, where the clearance between the workpiece and the chuck is only 0.5 millimeters; the approach covers motion generation, multi-view perception, historical memory, and contact feedback.
Enabot releases a family companion robot buying guide comparing EBO SE 2/EBO Mini Sport, EBO Air 2 Plus, and EBO Max across three price tiers.
Episode 170 of the RLChina Paper Seminar will be streamed live on September 30, 2026, from 14:00 to 14:30. Ma Shaojin, a master's student at Tianjin University, will present the paper "One Act, Many Masks: Effect-Marginal Policy Optimization for Long-Horizon Agents," with Guo Wei, a doctoral student at Tianjin University, moderating.
Kids can use robot pets to play interactive games, make up stories, ask questions to learn, and talk about their day and feelings. They can also practice simple responsibility through care tasks such as charging and putting the pet away. Research shows that social robots for children aged 6–10 can encourage kids to express emotions through direct questioning during emotional learning activities, and other studies have examined collaborative storytelling and vocabulary practice between children and social robots. Robot pets should supplement, not replace, real pets, peer companionship, and adult guidance.
Our robots are not driven by a single AI model. This is a deliberate choice. The first installment in our 10-part Physical AI series: where AI holds up in production, where it does not, and why safety sits outside the learning loop. https://www.atirobotics.ai/resources/educational-content/where-we-stand-on-ai/
Researchers are developing supernumerary robotic limbs (SRLs) that can be attached to the human body for human movement augmentation, unlike prosthetics that replace missing limbs. Existing SRLs are mostly experimental, controlled mainly by foot pedals, and brain-control approaches have low success rates. Research shows that with feedback provided by vibration motors and electrodes, people can learn the basic operation of controlling a third arm or a sixth finger in under an hour.
Boston Dynamics mechanical engineer Diane Heinle discusses durability testing for the Stretch robot: she is responsible for the dynamic cable design on the Stretch arm, cycling cables millions of times in the lab at speeds and forces far beyond real-world conditions, simulating years of warehouse wear in a matter of weeks. The durability lab is known as the "Tick-Tock Room" because the cable testers swing back and forth like metronomes, and the test data is also used to refine the design and maintenance schedules, avoiding downtime at customer sites.
In the first installment of its "Why" series, ROBROS explains why Whole Body Control (WBC) is key to keeping humanoid robots balanced: movements such as extending an arm, changing posture, or taking a step all shift the body's center of mass, requiring a coordinated whole-body response.
A team led by Assistant Professor Zhang Shuyang at Shenzhen University's College of Mechatronics and Control Engineering used the PaXini PX6AX multi-dimensional tactile sensor as its validation platform to address cross-operator generalization in tactile material recognition, proposing the lightweight tactile model Lite-TacFormer, which improves cross-operator generalization through a few-shot calibration mechanism and features an extremely small parameter count, minimal sample dependence, and a very low computing threshold. Built around the self-developed FUSE tactile sensing chip, the PX6AX maintains millinewton-level accuracy and microsecond-level response after tens of millions of force measurements, and its integrated sensing-computing-transmission design shortens experiment cycles.
Recording of the "Basics" chapter videos for the D-Robotics RDK Mini Class is essentially complete, and they have been published on Bilibili's "D-Robotics Lab" and YouTube's "RDKLabDev" for developers getting started with the RDK X5 / S100 development boards.