The research was presented in Frontiers in Robotics and Artificial Intelligence in 2026 and evaluated both an artificial-intelligence perception system and its integration with a physical arm. The machine is designed for household-assistance applications, in which a user might ask a robot to retrieve one particular item from a table surrounded by clutter. Its arm weighs 2.2 kg, has a payload capacity of 1.5 kg and ends in a one-degree-of-freedom gripper that closes its thumb and four remaining fingers. The arm comprises three coupled tendon-driven modules: a three-degree-of-freedom shoulder, a two-degree-of-freedom elbow and a two-degree-of-freedom wrist. Its lightweight, elastic tendons were chosen partly to suit operation close to people.

To perceive the scene, the researchers converted an RGB-D image into a three-dimensional point cloud and used only each point’s spatial coordinates, excluding color information to reduce sensitivity to texture and lighting. A shared backbone extracted features for two linked tasks: estimating full six-degree-of-freedom grasp poses and assigning points to semantic object categories. The latter task links each possible grasp to a recognized object, allowing a user to select a target category in a graphical interface. The system then executes the highest-scoring grasp associated with that category, defined by the candidate point, approach direction, rotation around the approach axis, gripper width and gripper depth.

The central network innovation was a 3D Mixture-of-Experts UNet based on the MinkUNet architecture. Rather than activating an entire enlarged model, the design used eight parallel expert networks and a gating mechanism that selected the top two for each input. The experts replaced convolutional layers in the two deepest residual blocks of the encoder, where features represent broad object geometry and scene context. An auxiliary load-balancing loss discouraged the gate from routing nearly all inputs to a small subset of experts. Evaluation showed that the eight experts were selected with broadly balanced frequency, indicating that expert collapse did not occur, although tests of alternative expert counts, routing arrangements and block placements found no configuration that consistently dominated the chosen design.

The multi-task model contained 24.2 million parameters in total, of which 17.1 million were active during inference because only two experts ran at once. By comparison, a separate semantic-segmentation model using the study’s conventional 3D UNet backbone would have required at least 14.7 million additional active parameters. The model was trained for 40 epochs with a batch size of four on a single NVIDIA GeForce RTX 4080 SUPER graphics processor. It used cross-entropy losses for objectness and semantic classification, smooth L1 loss for continuous grasp quantities, and several task-specific terms for approach direction, gripper dimensions and grasp quality. Label smoothing was applied to segmentation with a coefficient of 0.1.

Benchmark testing used GraspNet-1Billion, a dataset of 97,280 RGB-D images from 190 cluttered real-world scenes containing 88 general objects and more than 1.1 billion grasp-pose annotations. The 190 scenes were divided into 100 training scenes and 90 test scenes, with the test scenes further separated into seen, similar and novel object groups. Because many of the objects had similar shapes, the researchers assigned the 88 objects to seven geometric categories: pump bottle, tube, container, irregular-shaped object, spherical object, elongated or cylindrical object, and cuboid object. Camera-specific train and test sets were evaluated separately for RealSense and Kinect data. The proposed backbone generally improved grasp estimation and segmentation over 3D UNet and PointNet++ alternatives, with gains reaching 5.24 average-precision points for grasp estimation and 3.42 mean intersection-over-union points for segmentation.

Those aggregate improvements did not apply uniformly. Compared with the 3D UNet backbone, the Mixture-of-Experts model improved RealSense grasp-pose average precision by 2.43 points on seen objects, 1.98 on similar objects and 2.14 on novel objects. For Kinect data, it improved seen and similar results by 5.24 and 2.25 points, respectively, but the novel-object result fell by 0.65 points. Segmentation similarly declined by 1.38 mIoU points on the RealSense novel subset, even as grasp-pose performance improved on that subset, suggesting a trade-off between the tasks’ shared representations. The authors associated weaker Kinect performance with sensor characteristics such as depth noise, missing points and differences in point-cloud density. On the similar and novel test sets, segmentation results were affected by systematic inter-class confusion: cuboid objects could be classified as containers, while pump bottles could be confused with tubes.

The arm itself used a geometry-based inverse-kinematics method to resolve its redundant seven-degree-of-freedom structure. When the end-effector pose is fixed, the elbow can occupy multiple positions, so the method introduced an arm angle and searched its permitted range in fixed increments. It selected the angle that minimized the sum of the absolute joint angles, producing a complete set of joint commands. The motion planner then interpolated position linearly and orientation with spherical linear interpolation. In the physical workflow, the end-effector first moved to a pre-grasp pose offset opposite the estimated approach vector, advanced to the selected grasp, lifted the object vertically, transferred it to a predefined placement location and released it.

The laboratory evaluation involved 10 target objects in scenes containing at least four objects each, with object numbers, positions and orientations varied manually between trials. Each target received 10 pick-and-place attempts, and success required grasping the specified object without toppling it, lifting and transporting it, then releasing it without a drop. All three seen objects achieved 100% success. Among unseen objects, the individual rates ranged from 90% to zero: objects 4 and 6 each succeeded in nine trials, objects 7 and 9 in seven, object 5 in four, object 8 in one, and object 10 in none. The 52.9% average for unseen objects therefore concealed substantial differences associated with object geometry, size and material. The researchers identified several contributing constraints, including incomplete depth measurements from translucent plastic, limited contact options imposed by the one-degree-of-freedom gripper, and difficulty avoiding the supporting surface when few stable grasp regions existed.

The complete system, including the robotic-arm controller, used 5.1 GB of RAM and 3.12 GB of video memory and processed images at an average of 20 FPS in the experimental setup. These measurements suggest that the workload could fit on a compact edge computer such as a Jetson Orin2, but its performance on such hardware was not tested. The prototype also had important functional limits. Semantic categories allowed a user to request a broad object type, yet two instances of the same type could not be selected separately without instance-level segmentation. The study proposes future work on more adaptable multi-degree-of-freedom hands, improved depth acquisition for translucent objects, better mixture-of-experts tuning and instance segmentation. It also suggests voice or electromyography-based target selection for users with limited physical mobility, an application direction that was proposed rather than evaluated.

Subject of Research: A 7-DOF robotic system for autonomous perception and grasping of user-selected general objects

Article Title: Development of a 7-DOF robotic arm control system for autonomous grasping of user-specified general objects

Article References: Bai, S., Jiang, Y., Yokoi, H., & Togo, S. (2026). Development of a 7-DOF robotic arm control system for autonomous grasping of user-specified general objects. Frontiers in Robotics and AI, 13, 1955232. https://doi.org/10.3389/frobt.2026.1955232

Image Credits: AI Generated

DOI: 10.3389/frobt.2026.1955232

Keywords: robotic grasping, assistive robots, 7-DOF arm, point-cloud perception, grasp pose estimation, semantic segmentation, mixture of experts, inverse kinematics, tendon-driven robot, pick-and-place, Development, robotic

The research, conducted by Matthias Mayer and Matthias Althoff of the TUM School of Computation, Information and Technology, tackles what roboticists call base-pose optimization. A robot’s base pose is its position and orientation on the factory floor, and it determines which points in the workspace the arm can comfortably reach, how contorted its joints must become to hit each target, and how long every cycle of the production process takes. Moving the base by even a few tens of centimeters can shave seconds off a cycle time, and over millions of cycles those seconds translate into substantial money. Crucially, unlike buying a faster robot, repositioning the one you already have costs essentially nothing.

Despite decades of interest, the field had never established which optimization algorithm actually performs best. Previous studies each championed their own method on their own tasks, with no common benchmark to arbitrate between them. Mayer and Althoff set out to close this gap by implementing four families of optimizers on identical footing: exhaustive search approximated by uniform random sampling, genetic algorithms, Bayesian optimization, and, for the first time in this domain, stochastic gradient descent. They adapted Adam, the optimizer that underpins much of modern neural network training, to the robot placement problem, requiring only that the robot’s forward kinematics be well defined rather than any special wrist geometry assumed by earlier gradient-based work.

The technical setup is a two-level optimization. At the outer level, the chosen algorithm proposes candidate base poses within an allowed region, roughly a cubic meter around a nominal position with either fixed or unrestricted orientation. At the inner level, a task solver evaluates each proposal. Fast filters first reject poses farther from any goal than the robot’s maximum reach, a check that takes milliseconds, then reject poses lacking collision-free inverse kinematic solutions, which takes about a tenth of a second. Surviving candidates go to the RRT-Connect path planner and a time-optimal trajectory generator, which produce an executable motion through all required goal poses. The cycle time of that motion becomes the cost. For gradient descent, the researchers used automatic differentiation in PyTorch to compute how the cost changes with the base parameters, then applied Adam steps, restarting from random initial guesses to escape local minima.

The benchmark tasks came from CoBRA, a composable benchmark suite for robotics applications. The team built a set of 100 simple synthetic tasks with three cubic obstacles and three goal poses, then tested generalization on harder variants with five goals and obstacles, on tasks derived from 3D scans of real factories where robots feed CNC machines amid irregular geometry, and on deliberately perverse edge cases where all goals sit tightly against obstacles. Hyperparameters for the three tunable methods were optimized with Optuna using a tree-structured Parzen estimator, trained on the first 70 percent of the simple tasks and evaluated on the remainder plus the unseen sets. Every algorithm received an identical budget of 20 minutes per task, and results were aggregated over five random seeds with 95 percent confidence intervals and Wilcoxon signed-rank tests.

The headline finding is that stochastic gradient descent wins on reliability. Across all tasks, SGD achieved a significantly higher success rate than every competing method, and on the real-world, 3D-scanned factory tasks it solved more than 90 percent of them, beating random sampling, genetic algorithms, and Bayesian optimization alike. On the difficult synthetic set with five goals and five obstacles, SGD found valid placements in roughly 57 percent of position-plus-orientation runs, compared with about 44.8 percent for random sampling and 39.4 percent for Bayesian optimization. The advantage appeared most strongly exactly where deployments are hardest: cluttered environments and tasks demanding many distinct reach targets.

Genetic algorithms, by contrast, claimed the prize for quality. Among the tasks each method managed to solve, GA delivered the lowest final cycle times on average, significantly undercutting both random sampling and Bayesian optimization across the full task suite. This edge was most pronounced in the edge-case set, where feasible base placements are squeezed into narrow zones; there, GA achieved an average best cycle time of about 14.3 seconds versus more than 17.5 seconds for Bayesian optimization. Bayesian optimization, often a darling of expensive black-box problems, fared worst overall, showing the lowest success rate on the constrained edge tasks when orientation was fixed, around 5.4 percent, and significantly higher costs than all other methods across the board.

The study also yielded practical lessons about the search space itself. Allowing the optimizer to choose the robot’s orientation as well as its position, rather than fixing the base facing, increased success rates on the hardest edge tasks for every algorithm, sometimes spectacularly: random sampling jumped from about 14.6 percent to 54.4 percent, and Bayesian optimization from 5.4 percent to 25 percent. The authors note this suggests that the extra mounting effort of a freely oriented base pays off for complex tasks. Generalization, however, came at a cost for everyone. Success rates dropped when methods tuned on simple tasks faced harder distributions, yet all algorithms still exceeded 80 percent success on the realistic factory scans, with SGD leading at better than 90 percent.

The contrast with the baseline is stark: a dummy optimizer that simply kept the nominal task pose solved only 47 percent of simple tasks, 18 percent of hard tasks, and not a single task from the real or edge sets. In other words, naively placing a robot where the task description happens to put it is frequently a recipe for failure or slow cycles, while even modest optimization effort pays for itself within a 20-minute compute budget. Because cycle time correlates positively with other costs such as energy consumption, faster cycles likely mean greener ones too.

Perhaps most importantly for the field, the entire comparison is now open. All benchmarks, implemented methods, tuned hyperparameters, and generated solutions are publicly available through the CoBRA platform and the team’s code repository, giving future researchers a standardized baseline against which novel placement algorithms can be measured. As robots grow lighter and more modular, requiring less sturdy foundations and inviting more frequent relocation, the ability to compute an optimal parking spot in minutes rather than engineer one by hand could become a standard step in every deployment. This study turns an intuitive art into a benchmarked science, and it suggests the smartest placement algorithms may be the same ones that taught machines to see and speak.

Subject of Research: Comparison of optimization algorithms for industrial robot base-pose placement

Article Title: Smart placement, faster robots—a comparison of algorithms for robot base-pose optimization

Article References: Mayer, M., & Althoff, M. (2026). Smart placement, faster robots—a comparison of algorithms for robot base-pose optimization. Frontiers in Manufacturing Technology, 5, Article 1642524. https://doi.org/10.3389/fmtec.2025.1642524

Image Credits: AI Generated

DOI: 10.3389/fmtec.2025.1642524

Keywords: industrial robots, base-pose optimization, robot placement, stochastic gradient descent, Adam optimizer, genetic algorithms, Bayesian optimization, benchmarking, manufacturing automation, inverse kinematics, cycle time, CoBRA

The core insight behind Tiny-STM is that object permanence, the cognitive ability to understand that things continue to exist when hidden from view, does not require a learned memory mechanism of the kind used in large transformer-based systems. Instead, the researchers, Swarnajit Bhattacharya of National Yang Ming Chiao Tung University in Taiwan and Ian Chao of the University of Washington, paired a standard four-gate Long Short-Term Memory network with a deterministic four-slot coordinate bank. Each slot stores the most recently observed three-dimensional position of one task object. When a cube disappears behind a stack of other blocks, its stored coordinate simply holds steady until the camera sees it again, a scheme the authors describe as a zero-order-hold update rule.

The architecture itself is deliberately conventional. The LSTM takes a seven-dimensional input vector combining the vector from the gripper tip to the target object’s stored position, the current tip position, and a one-dimensional infrared range reading from a Sharp GP2Y0A41SK0F sensor mounted on the gripper. A hidden state of 128 dimensions feeds a linear output layer that predicts an incremental three-dimensional displacement of the end effector. With 70,019 parameters in total, the network is tiny by modern standards, and after INT8 quantization through the TensorFlow Lite Micro pipeline it occupies just 68.4 KB, a small fraction of the ESP32’s 520 KB of static RAM.

Training relied on remarkably little data. The authors manually guided a Waveshare 5-DOF RoArm-M3 robotic arm through exactly 30 pick-and-place demonstrations, recording joint angles, cube centroids from a side-mounted camera, and infrared range measurements at each control step. Backpropagation through time with mean-squared-error loss and the Adam optimizer at a learning rate of 0.001 produced the deployed checkpoint. This data efficiency stands in sharp contrast to imitation-learning platforms such as ALOHA, which achieve 80 to 90 percent success on fine-grained bimanual tasks but require roughly ten minutes of teleoperated demonstrations and substantially more powerful compute.

The division of labor between hardware components is a key part of the design. A host PC or single-board computer captures camera frames and performs deterministic color-thresholding segmentation, converting detected pixel centroids into three-dimensional workspace coordinates through a calibrated pinhole camera model. The camera was calibrated with 25 checkerboard images using Zhang’s method, yielding a mean reprojection error of 0.3 pixels, while camera-to-base extrinsic calibration over 12 known points achieved a root-mean-square alignment error of 1.18 millimeters. The ESP32 then handles everything on the control side: the recurrent forward pass, analytical inverse kinematics, and PWM servo command generation, all in a measured 2.80 milliseconds.

The inverse kinematics routine reduces the five-degree-of-freedom arm problem to a planar three-degree-of-freedom calculation, exploiting the fact that the task uses a top-down grasp with a vertically oriented gripper. Base rotation comes from an arctangent of the target coordinates, while the shoulder and elbow angles follow from two-link geometry via the law of cosines. The authors handle three degenerate cases explicitly: base singularity when the target lies on the rotation axis, arm singularity near full extension where the Jacobian becomes ill-conditioned, and unreachable targets outside the geometric annulus defined by the link lengths, which are simply rejected so the arm retains its previous pose.

Physical evaluation consisted of 20 trials in which the arm sorted four labeled cubes from randomized positions into an ordered arrangement, including a vertical stacking sequence in which the fourth block becomes completely hidden behind the first three before its final grasp. The single deployed INT8 checkpoint completed 17 of 20 trials, an observed success rate of 85.0 percent with a two-sided 95 percent exact binomial confidence interval of 62.1 to 96.8 percent. The authors are careful to note that this interval describes uncertainty in the 20 observed outcomes of one checkpoint, not variability across independently trained models, since the original training seed and stopping rule were not retained.

The three failures are instructive. Mechanical backlash in the servo gearboxes introduced joint-level positional uncertainty of roughly 1.5 to 2.5 millimeters, which can accumulate across joints. Fixed color thresholds proved vulnerable to strong lighting changes and shadows that shifted pixel values out of range, degrading the segmentation masks. And when an occluded object was physically moved while hidden, the stored coordinate became stale, causing the controller to reach toward a position where the object no longer was; the gripper tolerance is approximately 18 millimeters. The authors formalize this limitation in a corollary showing that if an occluded object is displaced, the memory error grows at least in proportion to the displacement.

On latency, the measured ESP32 computation of 2.80 milliseconds combines with the nominal 33.3-millisecond camera period at 30 frames per second, profiled segmentation and Wi-Fi communication times, and a datasheet-based actuator response allowance to yield an aggregated end-to-end latency estimate of 92.6 milliseconds, equivalent to roughly 10.8 hertz. The authors emphasize that this figure is an estimate assembled from components rather than a single timed end-to-end measurement. Peak SRAM usage on the microcontroller was 94.8 KB, or 18.23 percent of capacity, leaving substantial headroom. Offline comparisons suggest the FP32 model would occupy 273.5 KB with an estimated 11.20 milliseconds of inference, underscoring the value of INT8 quantization for embedded deployment.

The study is notable as much for its honesty about boundaries as for its results. The authors explicitly state that no ablation was performed to separate the contribution of the coordinate memory from that of the recurrent state, that no numerical comparison with RT-2, OpenVLA, TinyVLA, or Diffusion Policy is possible given different tasks and hardware, and that the color-thresholding front end was never validated against ground-truth masks. Future work, they write, should include matched ablation studies with multiple training seeds, documented reproducibility protocols, perception validation, tests of behavior when occluded objects are displaced, and scaling experiments beyond four tracked objects. Even so, Tiny-STM offers a compelling demonstration that a form of object permanence can live in 68 kilobytes, bringing occlusion-aware manipulation within reach of the cheapest embedded hardware.

Subject of Research: Lightweight recurrent neural network controllers with external coordinate memory for occlusion-aware robotic manipulation on microcontrollers

Article Title: Lightweight LSTM-based memory architecture for occlusion-aware robotic manipulation on microcontrollers

Article References: Bhattacharya, S., & Chao, I. (2026). Lightweight LSTM-based memory architecture for occlusion-aware robotic manipulation on microcontrollers. Discover Informatics, 1(1), Article 23. https://doi.org/10.1007/s44564-026-00023-0

Image Credits: AI Generated

DOI: 10.1007/s44564-026-00023-0

Keywords: TinyML, LSTM, robotic manipulation, object permanence, occlusion, ESP32, microcontroller, imitation learning, inverse kinematics, INT8 quantization, embedded robotics, recurrent neural networks