Dexterous Hand in Practice 08 focuses on single-joint control, explaining how to implement position PID closed-loop control so that a single joint moves stably according to commands. It is part of a hands-on tutorial series aimed at the implementation process of joint-level control for dexterous hands.
Part 06 of the Dexterous Hand in Practice series discusses how STM32, ESP32, and Raspberry Pi divide responsibilities in dexterous hand control. The content centers on the roles each of these three types of control boards plays, and it is a tutorial-style article in this hands-on series.
HAMPER reproduces Unitree G1 basket carrying in Isaac Lab, using a PPO whole-body policy and a five-stage state machine to chain together approach, squatting, threading the handle, and lifting and carrying, with pretrained weights provided for real replay on a local machine.
dex-retargeting analyzes motion retargeting from human hands to robotic dexterous hands, using a Huber loss on wrist-to-fingertip bone vectors as the objective for per-frame numerical optimization. DexPilot builds on this method by adding grasp detection, projected targets, and temporal low-pass filtering, with the Allegro and Shadow hands converging per frame at millisecond-level speed.
This practical guide to Zephyr stepper motor control covers three features: microstepping, positioning, and event callbacks. It explains how to use microstep driving, position control, and the event callback mechanism for stepper motors.
REEMAN autonomous forklifts use laser SLAM navigation and can connect to existing WMS, MES, and ERP systems without magnetic strips or QR codes; the integration focus is on location mapping and open API connectivity rather than rebuilding navigation infrastructure. The guide covers site surveys, map creation, task triggering, and multi-vehicle fleet scaling. Each model is rated for 6 hours of continuous operation (fully loaded)/8 hours (empty), supports automatic return to charging, and some models allow manual battery swapping.
DeepBlue Embodied Intelligence examines the debate over embodied intelligence approaches along six main technical lines: modular planning, skill orchestration, imitation learning, reinforcement learning, VLA, and world models, drawing on 12 highly cited surveys from the past decade. The article notes that each approach answers the question of what robots should learn from, that the classic problems have not disappeared, and that what has changed is that the technical units for solving them are growing larger. It also mentions an industry estimate that general-purpose robots would need 100 million hours of data to reach a 70%–80% baseline success rate, and that more than 100 open-source embodied foundation models appeared worldwide in the past 18 months.
In October, the Xbotics Embodied Intelligence Community held three offline events: at the Shenzhen session, 48 developers used the OmniHand O10 dexterous hand to complete calibration and a crawling competition; at the Shanghai session, 38 developers completed the transition from MotrixLab simulation training to real-machine deployment of the Qinglong Nano within two days; and at the Hangzhou session, 22 people learned to replicate the MiniDuck little yellow duck.
In Lecture 7 of the "Twenty-One Lectures" series, the Xbotics Embodied Intelligence Lab converts the target object pose obtained in Lecture 6 into gripper actions, linking the grasp-and-place loop with five poses—pre-grasp, grasp, lift, place, and retreat—and a state machine.
Tianyi'an has launched a beginner CoppeliaSim robot simulation development course based on CoppeliaSim 4.9/4.10, teaching the software's basic operations through five projects: a six-link robot, a four-legged linkage robot, a line-tracking and collision smart car, a Mecanum-wheel car, and an E-puck smart car. The course is originally priced at 79.8 yuan, with a 2-person group rate of 59.8 yuan and a 3-person group rate of 39.8 yuan; after purchase, it can be viewed free for 2 years. Future updates are planned to cover Simulink/Python co-simulation, robotic arm control, and trajectory planning.
Using the Push-T tabletop pushing task as a vehicle, this piece fully reconstructs the mathematical formulation of action diffusion policies, covering cosine-schedule noise addition, the noise-prediction training objective, FiLM conditioning injection, and sliding-window execution. The author also uses public pretrained weights to complete a real inference evaluation across ten seeds.
For fine-tuning VLA models, the article presents a reproducible parameter-efficient training approach spanning from full SFT to LoRA. The content covers these two training paths for developers to reproduce according to the approach.
Gu Yue Ju analyzes the PlaCo humanoid gait pipeline: a linear inverted pendulum trajectory optimization using jerk as the decision variable keeps the ZMP inside the support polygon, cubic splines connect the swing foot, and weighted QP inverse kinematics maps the center of mass and both foot targets into 26-dimensional joint velocities, completing continuous turning walking.
This article reproduces the C-1N hexapod robot in MuJoCo, testing the inference performance of an open-loop tripod gait and four official pretrained policies. The author quantifies the triple coordination of hexapod walking across four dimensions: phase, support margin, contact load, and residual action magnitude.
OpenCR-MuJoCo presents a discretized simulation method for tendon-driven continuum robots in MuJoCo, modeling the continuous backbone as a rigid chain with elastic joints, where bending is produced by tendon length differences. The work uses SoRoSim reference data and the official demo to validate static accuracy, release oscillation, and closed-loop tracking.
This study takes a bicycle robot with a reaction pendulum as its subject, derives the roll dynamics and the speed-dependent instability mechanism, and uses proximal policy optimization (PPO) to train coordinated control of the steering, rear wheel, and reaction pendulum. Experiments are carried out in PyBullet: natural fall, training, balancing, navigation, and disturbance recovery, concluding with obstacle-avoidance training.
Gu Yue Ju analyzes the MuJoCo real-time predictive control framework, which uses a residual cost to unify task descriptions, with four planners—predictive sampling, gradient descent, iLQG, and hybrid ILQS—replanning at every step within the same loop. A Cartpole comparison experiment reports the convergence speed and per-step computational cost of the four.
The article addresses the backlash and lost motion problem in robot joint reducers by building a dual-inertia backlash transmission model described with a Play operator, and designing a direction-reversal state machine with rate-limited feedforward compensation. It quantitatively compares three control schemes through three sets of experiments and provides a Python simulation implementation path.
Using bridge crane anti-sway as the object, the article derives the composite pendulum dynamics formed by the trolley, a homogeneous suspension rod, and the payload, and presents the construction of a multi-stage scenario tree, non-anticipative input sharing, and an equally weighted expected objective. The authors conducted measurement-intermittent comparative experiments between nominal and robust controllers in the MuJoCo physics engine.
The article explains Ruckig's jerk-limited online trajectory generation, covering the seven-segment profile of the triple-integrator-chain time-optimal problem, the selection of synchronization instants across multiple axes, and the closed-form solution for braking when limits are exceeded. It also presents measured tracking and target-jump replanning results for a two-joint arm in MuJoCo.
Addressing the issue of loud thuds when a humanoid robot lands, the investigation points to the difference in choosing between foot force control and position control. The article presents the root-cause diagnosis process for this abnormal noise in the form of a troubleshooting record, for developers to compare against their own control schemes.
This SmolVLA tutorial demonstrates the complete workflow for running SO-100 on lerobot, showing that small models can handle such tasks. It focuses on how SmolVLA works with SO-100, covering the entire chain from environment setup to execution.
The article documents unsupported operator issues encountered when running VLA models on RK3588 and compiles them into a list and troubleshooting log. The material provides only the title and summary; specific operators, model versions, or performance data are not given.
To address the inability to mix CVC600 and CVC700 controllers in a scheduling system, Kollmorgen offers a software-level compatibility solution: update the CVC700 to firmware with the BC suffix (latest 2.8.12 BC) and add the corresponding license in the scheduling system, enabling unified scheduling of newly added CVC700 vehicles within existing CVC600 legacy projects.
The 31 failures logged in the week before the competition, after deduplication, fall into 5 root cause categories: open-loop control, changes not taking effect, dead code, false feedback, and no attribution. Each root cause comes with real code and arithmetic reasoning, plus three actions you can follow directly.
The author uses EvergreenSLAM, a self-developed 2D lidar SLAM project, as Claude Code's spatial memory. Through the egs command line, each location corresponds to a folder and notes.md on the lidar map, with the agent writing notes itself, and a two-day real-world demonstration was completed on an XLeRobot.
Based on a survey of more than 20 mainstream quadruped robot projects on GitHub, the article maps the open-source quadruped technology landscape across five schools: kinematic control, full-stack engineering, optimization algorithms, reinforcement learning, and hardware platforms.
The fifth installment of the Gazebo Mecanum-Wheel Robot Simulation series discusses deliberately placing furniture in indoor scenes that LiDAR cannot scan, because mapping requirements conflict with exam requirements; the solution is to split this into two layers, a "permanent layer" and a "blind-spot layer."
A developer built a small desktop robot called Little ESP32 deskbot using a Waveshare ESP32-S3 1.46-inch OLED board paired with GH-S37 micro servos. They connected it to the free XiaoZhi AI service, flashed firmware using its sample code, and added their own scripts to implement facial expressions and servo movements.
The author completed Fetch, a robotics STEM project for beginner and intermediate learners costing about 300 USD. It uses a binary-classification object detection model trained with Edge Impulse and is the author's first autonomous robot.
In each round, Dijkstra selects the node with the smallest cumulative cost from the source among the nodes not yet finalized and formally finalizes it, then performs relaxation updates on its neighbors; under the premise of non-negative edge weights, this guarantees the globally shortest path. If negative-weight edges exist, a finalized node may be overturned by a cheaper route, so standard Dijkstra no longer guarantees correctness, and algorithms such as Bellman-Ford must be used instead.
Using Unitree H1, Isaac Sim 2024.1, IsaacLab, and PPO as its stack, the article walks through a complete sim-to-real pipeline: converting URDF to USD, scene setup, domain randomization configuration, PPO training, ONNX export, and ROS 2 deployment on the real robot.
REEMAN outlines an approach for retrofitting existing stretch wrappers using Ironhide autonomous forklifts: the forklift handles pallet pickup, placement, and warehouse transport, while the stretch wrapper continues to perform packaging, requiring a machine control interface, status feedback, and safety interlocks. In its wireless docking demonstration, pickup alignment deviation was within 2 cm and turntable placement deviation was within 5 cm, with the demonstration process taking about 6 minutes. There is also an Ironhide 3.0 deployment at a factory in Jingmen, Hubei; the two applications are independent of each other, and the data comes from REEMAN application materials and has not been independently verified.
A weekend PoC mounts an old Redmi 9 on a Thymio II using Lego, with the phone running ML Kit face detection (including smile probability) and image annotation, rough lighting/motion estimation, animated expressions, and voice output, while the robot only executes wheel-speed targets and sends back sensor data. Behavior is a small state machine: cruise → back up → turn → observe → stop steadily; when a face is detected it enters a "meet" state and can answer "what do you see." Obstacle-avoidance learning is simple: after a collision it tries turning for 1.2s, 2.4s, or 3.8s, scored by the free-driving time before the next contact.
REEMAN demonstrated a library inventory robot built on the Hot Wheels robot chassis, completing a mobility demo in indoor bookshelf aisles, but did not provide OCR recognition accuracy or inventory results. The chassis uses Laser SLAM and V-SLAM for mapping and localization, combined with lidar, 3D vision, and multi-sensor fusion for obstacle avoidance; published specifications include a 270° laser scanning angle and a maximum detection range of 25 meters.
The team recreated a real self-service tea station as a physically interactive 3D environment in Isaac Sim, placed a Unitree G1 in it, and used a script to execute the five-step tea-making motion. The robot's motions were not extracted from human video; the entire run took place in simulation with physics enabled. The G1 still shows imperfect contact and slipping when interacting with objects, and there is currently no learned policy or automatic trajectory generation. The team is exploring whether such environments can serve as a source of synthetic robot demonstration data, including for robots with different embodiments.
On the Quaddle mini robot dog's ESP32-S3 firmware, six Claude Code prompts were used to complete a skull eyes effect, from sketch and choosing a design to preview and on-device display. The firmware will be open-sourced in early December, and the team is also collecting sketches or images for the next Halloween expression.
Someone used Claude Code to vibecode directly in the OpenCat ESP32-S3 firmware of the Quaddle mini robot dog, creating a Halloween skull-eyes expression in about 12 minutes through six prompts. The robot dog's open-source ESP32-S3 firmware will be released in early December, at which point others will be able to reproduce the same process.
The author debugged and fine-tuned a motion controller based on the MPC/WBC framework for a self-developed quadruped robot, and used depth-camera images to build height maps. The robot can now climb stairs more reliably, step over obstacles up to 11 cm high (with 18 cm ground clearance), and handle soft ground with greater confidence. The controller is written in C++, is ROS2-compatible, and has been open-sourced in the voltdog/mors_quadruped repository on GitHub. Next, the author will focus on improving software reliability, including automated device connection tests and standing up from any posture.
Following the thread of "how to solve problems in different environments," the article reviews five categories of path planning paradigms in embodied intelligence: graph search (Dijkstra, A*, Hybrid A*), sampling methods (RRT, PRM), artificial potential field (APF), optimal control (MPC), and reinforcement learning and data-driven methods.