COBOTA 2026 Abstracts

Area 1 - Bridging the Gap in COllaborative roBOtics: from Theory to real Applications
Full Papers Short Papers

Area 1 - Bridging the Gap in COllaborative roBOtics: from Theory to real Applications

Full Papers
Paper Nr: 6
Title:

Cobot-Based Operator Assistance for Manual Meat Processing: Experimental Validation on Porcine Cuts
Authors:

Alexis Babut, Chedli Bouzgarrou, Laurent Sabourin and Nicolas Bouton
Abstract: Manual meat processing operations — muscle separation, anatomical trimming and deboning — impose severe musculoskeletal loads on skilled operators, yet the spatial heterogeneity of biological tissue makes them resistant to full automation. This paper presents the experimental validation of a cobot-based operatorassistance system in which a UR30 collaborative robot carries an instrumented boning knife fitted with two six-axis force-torque sensors: one isolating the operator’s guiding wrench at the handle, the other capturing the aggregate tool load at the robot flange. An estimation stage jointly separates the operator interaction wrench from the blade–tissue reaction wrench in real time at 500 Hz, and a Fractional Shared Compliant (FSC) admittance law drives the robot from the operator’s guiding force whilst returning, on the selected cutting axes, a tunable fraction α of the measured tissue reaction to the operator’s hand. Unlike teleoperation, where force feedback must be artificially rendered by a master device, human, robot and task are here in direct physical series contact, and the controller regulates how much of the task reaction is felt at the handle. The system is validated on real porcine loin and pork belly cuts with an experienced butchery operator performing continuous trimming passes and deboning passes along the ribs. Recorded operator and cutting wrenches demonstrate that the robot sustains the dominant share of the tissue resistance whilst the operator’s guiding effort remains low and commensurate with intentional gesture, establishing the feasibility of fractional effort sharing for softtissue cutting under realistic processing conditions.
Download
Paper Nr: 8
Title:

Facial Emotion Recognition Based on FAUs and DINOv3 Embeddings with Reliability-Gated Fusion
Authors:

Giuseppe Lorenzo Di Prima, Francesco Termine, Ankit Kumar, Giuseppe Sutera, Filippo Sanfilippo and Mario Collotta
Abstract: Facial Emotion Recognition is essential for natural and effective Human–Robot Interaction. Facial expressions can be represented through interpretable geometric descriptors, such as Facial Action Units, or through embeddings extracted by large pre-trained models, which generally provide richer but less interpretable representations. This work proposes Reliability–Gated Fusion (RGF), which combines these complementary paradigms through per-sample reliability estimates. Each stream produces a categorical distribution and a calibrated reliability score, while a learned gate dynamically weights their contributions, providing an interpretable fusion mechanism through directly inspectable gate weights. A systematic comparison is conducted against three face-representation paradigms and two recently proposed multimodal affect-recognition frameworks: i) interpretable Face Action Units only; ii) image embeddings only, obtained through a frozen DINOv3 Vision Transformer; iii) video-dynamics embeddings, obtained through a frozen V-JEPA 2 video transformer; iv) Adaptive Entropy-Gated Contrastive Fusion; and v) Adaptive Gated Fusion Network. Under the same experimental protocol, RGF achieves 70.65±2.74% accuracy, 69.87±2.39% macro F1-Score and 72.14±2.27% UAR across three independent runs, outperforming all evaluated competitors and achieving a 0.56 percentagepoint accuracy gain over the geometric-only approach, which represents the best-performing baseline.
Download
Paper Nr: 9
Title:

Keeping the Franka Emika Panda Alive: A ROS 2 Stack with a Reliable Position Interface
Authors:

Antonio Langella, Davide Risi, Vincenzo Petrone, Enrico Ferrentino and Pasquale Chiacchio
Abstract: This paper presents an open-source software stack that restores ROS 2 support for the Franka Emika Panda robot while resolving the long-standing unreliability of its external position control interface. We first analyze the root causes of unstable position control and show that the observed vibrations and protective stops arise from the timing of the external control loop and sampling jitter, rather than from limitations of the robot itself. Building on this analysis, we introduce an asynchronous hardware interface that decouples real-time communication from the ROS 2 control loop, a rate-matching mechanism for slower command sources, and a position-domain reference generation strategy that produces reliable, smooth position commands. Experimental validation shows that the proposed architecture reliably tracks velocity references by reducing motion artifacts introduced by the official implementation, and the stack is validated across motion planning, compliance control, position-controlled manipulation, and haptic teleoperation on two independent Panda platforms. By restoring a modern, reliable, and open ROS 2 ecosystem for the Panda, this work lowers the barrier to developing safe, responsive, and reproducible human-robot collaboration applications that integrate planning, perception, interaction, and shared autonomy. Code and videos are available on our website at https://sites.google.com/unisa.it/fer-ros2.
Download
Paper Nr: 10
Title:

Robust Learning from Demonstration Pipeline Based on Inertial Motion Capture and Force-Synchronised Demonstration Capture for Industrial Demoulding Tasks by Collaborative Robots
Authors:

Alejandro Solar Iglesias, Gabriel Novas Domínguez, Saltanat Seitzhan and Juan Antonio Corrales Ramón
Abstract: The main objective is to develop and validate a pipeline capable of transferring human movements to a collaborative robotic arm (UR5e) through probabilistic trajectory modelling and Learning from Demonstration (LfD). The methodology integrates data acquisition from Xsens inertial sensors and force/torque sensors, skeletal modelling in OpenSim, and kinematic calibration in the MuJoCo physics simulation environment interfaced through ROS 2 middleware. Gaussian Mixture Models (GMM) and Gaussian Mixture Regression (GMR) are used to model the overall behaviour and its spatial variability. One of the main contributions is the development of a statistical preprocessing module based on the Interquartile Range (IQR), which automatically detects and excludes erroneous demonstrations caused by tracking or calibration failures, improving model accuracy, prediction interval width (MPIW), and dynamic smoothness (Jerk). The pipeline is validated in simulation and deployed on the physical robot through open-loop trajectory reproduction with offline force monitoring. Seven configurations of the number of Gaussian components K are evaluated, and K = 8 is selected after validation on the digital twin and on the real UR5e, obtaining an RMSE of 11.66 mm, PICP coverage of 100%, and zero safety alarms during execution.
Download
Short Papers
Paper Nr: 5
Title:

Automated IMU-Based Ergonomic Evaluation and Postural Risk Assessment for Industrial Automation Frameworks
Authors:

Jorge Galindo, Alejandro Solar, Juan Antonio Corrales Ramón and Hernan Abaunza
Abstract: This research presents an automated framework for the ergonomic evaluation of repetitive manual work routines in industrial environments, aiming to identify musculoskeletal injury risks and establish a deterministic data baseline for future robotic automation in order to reduce it. Motion capture is achieved using a network of wearable Inertial Measurement Units (IMUs) equipped with triaxial accelerometers, gyroscopes, and magnetometers. This methodology provides high precision, repeatability, and environmental immunity against lighting variations and occlusions common in computer vision or electromyography, especially in uncontrolled industrial workplaces. The acquired kinematic data is mapped onto an unactuated upper-body musculoskeletal model within an open-source biomechanical simulation environment, and relative joint angles are computed through optimization-based inverse kinematics. This framework digitizes and automates the continuous calculation of the Loading on the Upper Body Assessment (LUBA) ergonomic index via a custom computational pipeline, replacing traditional frame-by-frame manual observations. The experimental validation across sequential task iterations demonstrates a significant reduction in both the peak Postural Load Index (PLI) and the cumulative time exposure within critical risk thresholds, validating the computational pipeline as a robust tool for proactive industrial workstation design and human-to-robot task mapping.
Download
Paper Nr: 7
Title:

Transformer-Based Ego-Motion Filtering for Tracking in Eye-in-Hand Event-Driven Visual Servoing
Authors:

Gabriel Barrigón-Gómez, Adrián Fernández-Casas, Carlos Mateo-Agulló, Juan Antonio Corrales Ramón and Cédric Demonceaux
Abstract: Event cameras provide asynchronous and low-latency visual measurements that are well suited to dynamic robotic applications. However, their use in eye-in-hand visual servoing remains challenging because robot-induced camera motion generates events that may obscure the motion of the target object. This paper presents an event-based image-based visual servoing framework for robotic manipulation. Events acquired within temporal windows are accumulated into image-like representations and processed using semantic segmentation to separate ego-motion-related events from those associated with the target object. The resulting object-motion features are used to define the visual error of a velocity-based IBVS controller. The approach is evaluated through two simulated sequences and experiments on a real robotic platform involving the tracking of a moving object. The results show that SegFormer provides the best overall trade-off between segmentation accuracy and inference speed.
Download
Paper Nr: 12
Title:

Slice More, Push Less: Admittance-Controlled Force Amplification for Collaborative Meat Cutting
Authors:

Saltanat Seitzhan and Juan Antonio Corrales Ramon
Abstract: This work presents the experimental validation of a force amplification strategy for human-robot collaborative meat cutting. The system employs a dual-sensor instrumented knife end-effector that independently captures operator-applied forces at the handle and cutting contact forces at the blade. An admittance controller reads the operator’s intent from the handle sensor and generates robot velocity commands, while a programmable gain selectively amplifies the lateral (slicing) component of the human force input, with the robot contributing the remaining cutting effort. Experiments on lamb using a UR5e robot arm with gain factors of 1.6 and 2.0 demonstrate consistent collaborative cutting, with the dual-sensor setup enabling quantitative decomposition of the shared effort between operator and robot. To assess whether lateral amplification is mechanically justified, we apply Atkins’ fracture-mechanics model for slice/push cutting and compute the slice/push ratio ξ directly from blade sensor data during active cutting. Measured values of ξ ≈ 2.4–2.7 (median during active cutting) place the system well into the slicing-dominated regime, where the Atkins model predicts a 62–66% reduction in total cutting force relative to pure normal-force pressing, providing an independent theoretical validation of the amplification strategy.
Download
Paper Nr: 13
Title:

Design of a Low-Cost Planar Delta Robot for Agri-Food Sorting Applications
Authors:

Simão Lopes, Martim Aguiar, Yebo Lu and Pedro Dinis Gaspar
Abstract: Agri-food sorting lines require fast, repeatable and cost-effective robotic systems capable of handling variable produce without mechanical damage and without compromising throughput. This paper presents the design and preliminary validation of a low-cost 2-DOF planar Delta robot intended for the lateral exclusion of defective peaches from a conveyor in a sorting line. The proposed architecture reduces mechanical complexity relative to 6-DOF robotic arms by restricting motion to the planar trajectory required for fruit rejection. The design methodology combines parametric geometric modelling, analytical inverse kinematics, force analysis in the critical position, electromechanical component sizing, and Design for Additive Manufacturing of custom supports and connection elements. A NEMA 23 stepper-motor-based actuation system, belt–pulley transmission, linear guides, limit switches, and an Arduino Mega 2560 control architecture were selected and assembled for subsystem validation. Preliminary tests confirmed the correct operation of the electrical architecture, motor-driver communication, homing routine, and speed variation between 200 and 2000 steps s⁻¹. The results support the feasibility of the proposed architecture as a low-cost actuation platform for future integration with soft grippers and computer vision in agri-food sorting applications.
Download