Frontiers in Robotics and AI

Design and evolution of a triad twisted string actuator for controlling a two degrees of freedom joint: improving performance and simulating active transmission adjustment

2026-04-01

Damian Crosby, Joaquin Carrasco, William Heath, Lutong Li, Andrew Weightman

Actuated universal joints are used in a wide range of robotic applications, including mobile snake robots, snake-arm robots and robotic tails. They are employed in applications such as search and rescue and confined space inspection. These can use remote cables, fluid driven systems, or inline motors. To realise the benefits of inline actuation while keeping the system compact with a high power to weight ratio, an actuated universal joint (AUJ) was developed using an ‘‘antagonistic triad’’ of three twisted string actuators in our previous work. However, the design had numerous drawbacks in its prototype form, namely, a limited angle range, poor accuracy due to the angular feedback sensors used, and issues with string failure due to mechanical design choices. In this publication, we performed a root-cause analysis of these issues, and partially or fully mitigated some of them by reducing the distance between the twisted string actuator (TSA), removing geometry which caused premature string failure, and exchanging the angular feedback sensors for more accurate ones. As a result, angle range was increased from ± 14.5° to ± 26° for a single axis, and ± 6° to ± 20° for a dual axis movement. Angular feedback sensor accuracy increased from ± 0.21° to ± 0.11°, and no string failures occurred within load limits. The performance of the mechanism was further characterised with additional experiments for increased follower load and angular velocity. A novel method to adjust the transmission ratio during operation (active transmission adjustment) was proposed and simulated, and its advantages over existing mechanisms for a snake robot in a multi-segment configuration were theoretically evaluated.

Editorial: Harnessing visual computing to revolutionize manufacturing efficiency and innovation

2026-03-30

Edmanuel Cruz, Francisco Gomez-Donoso, José Carlos Rangel, Cesar Pinzon-Acosta

FRMD: fast robot motion diffusion via trajectory-level consistency distillation

2026-03-25

Xirui Shi, Yi Hu, Jun Jin

Foundation models for embodied artificial intelligence (Embodied AI) increasingly adopt diffusion modules as the action generation core of vision–language–action (VLA) policies, but the diffusion module’s iterative denoising imposes prohibitive inference latency for real-time deployment. We address this bottleneck in isolation by rethinking the diffusion action generation module itself. We present Fast Robot Motion Diffusion (FRMD), a fast robot motion diffusion framework that (i) operates in trajectory-parameter space by predicting movement-primitive coefficients in a low-dimensional manifold, and (ii) collapses multi-step sampling into a single inference step via trajectory-level consistency distillation over the probability-flow ordinary differential equation (ODE). Concretely, FRMD replaces stepwise action generation with a one-pass mapping from noise to full trajectories, followed by a fixed-cost basis expansion; this reduces policy latency from hundreds to tens of milliseconds without modifying upstream vision or language encoders. On standard robotic manipulation task benchmarks, FRMD attains 7 times faster than the vanilla diffusion policy and 10 times faster than the state-of-the-art MPD method, while matching the task success of multi-step diffusion policies. By targeting the diffusion component used throughout VLA systems, FRMD provides a plug-in, latency-optimized motion generator that preserves the advantages of diffusion and makes real-time embodied AI feasible.

Automated capture and transfer of human facial expressions to humanoid robots for realistic patient simulation

2026-03-24

Patricia Schwarz, Sebastian Spanknebel, Diana Immel, Rene Hurlemann, Andreas Hein, Sandra Hellmers

Realistic reproduction of human facial expressions is essential for realistic interactions between humans and humanoid robots. This work presents a data-driven framework for transferring human facial expressions to a humanoid robot and a virtual avatar, aiming to enhance emotional expressiveness and assess its applicability in psychiatric training scenarios. The proposed approach enables cross-domain facial expression mapping while accounting for mechanical constraints of robotic actuation. A user study (n = 40) evaluated emotion recognition across three stimulus categories: human faces (H), unconstrained virtual avatars (A) and humanoid robots with limited facial actuation (R). Participants identified emotions from static images and from dynamic expression sequences, presented with and without speech. Perceived realism and uncanny valley effects were assessed using an eight-item questionnaire rated on a 7-point Likert scale. Results indicate that human-to-robot facial expression transfer is feasible but constrained by mechanical expressivity. Highly expressive emotions such as surprise (H: 87.5%; A: 57.5%; R: 65%) and fear (H: 45%; A: 27.5%; R: 57.5%) achieved moderate recognition rates, whereas subtle emotions such as anger (H: 65%; A: 40%; R: 12.5%) and disgust (H: 60%; A: 10%; R: 22.5%) were poorly recognized on the robot. Dynamic expressions combined with speech significantly improved recognition. These findings demonstrate the feasibility of transferring human facial expressions to humanoid robots while highlighting current limitations of robotic facial actuation. The proposed framework provides a promising basis for emotionally realistic patient simulation and training applications in mental healthcare.

Informing robot design through early public engagement: lay perceptions of soft versus rigid socially assistive and rescue robots

2026-03-19

J. Fenn, L. Estadieu, M. Gorki, I. Monno, F. Tauber, J. Teichmann, S. Levy-Tzedek, T. Speck, O. Müller, A. Kiesel

As soft robots become more prevalent in society, it becomes increasingly important to understand how laypersons evaluate their risks and benefits relative to conventional rigid robots. This article investigates public perceptions of soft versus rigid embodiments of socially assistive robots (SAR) and rescue robots (RR) and explores how these perceptions can inform early-stage robot design. We conducted an online study, using a scenario-based intervention design combined with Cognitive-Affective Maps (CAMs) to capture participants’ cognitive–emotional belief structures. In a first step, participants constructed CAMs depicting perceived risks and benefits of rigid SAR or RR. After reading a second scenario introducing the corresponding soft robot, they revised their maps, allowing a direct contrastive comparison between the first (rigid) and second (soft) scenario. Quantitative analyses showed that, across both application domains, post-intervention evaluations (after the soft-robot scenarios) were more positive than pre-intervention evaluations of rigid robots. Qualitative analyses revealed distinct argument structures: After learning about soft robots, participants added concepts emphasizing safety, emotional comfort, and adaptability, but also introduced concerns such as fragility and emotional dependence, whereas rigid robots were linked to precision, robustness, and efficiency, alongside worries about technical failure, data security, and emotional detachment. By integrating intervention-based CAMs with data-driven qualitative synthesis, the study demonstrates a scalable method for early public engagement that uncovers how laypersons qualitatively negotiate trade-offs between soft and rigid designs in plausible early-stage scenarios. These insights provide actionable input for human-centered design of soft robots, supporting responsible and socially aligned robot development.

Control flow graph based code optimization using graph neural networks

2026-03-11

Melih Peker, Ozcan Ozturk

Selecting a good set of optimization flags requires extensive effort and expert input. While most of the prior research considers using static, spatial, or dynamic features, some of the latest research directly applied deep neural networks to source code. We combined the static features, spatial features, and deep neural networks by representing source code as graphs and trained Graph Neural Network for automatically finding suitable optimization flags. We created a dataset of 12000 graphs using 256 optimization flag combinations on 47 benchmarks. We trained and tested our model using these benchmarks, and our results show that we can achieve a maximum of 48.6% speed-up compared to the case where all optimization flags are enabled.

Hip exoskeleton assistance with machine-learning-based state estimation improves gait kinematics of people with Parkinson’s disease

2026-03-09

Keaton L. Scherpereel, Jessica E. Bath, Anna Roumiantseva, Jacob Marks, Doris D. Wang, Patrick W. Franks

Exoskeleton assistance has the potential to address many gait related symptoms of Parkinson’s disease (PD). However, gait variability, a hallmark of PD, makes designing exoskeleton controllers uniquely challenging. We sought to overcome the challenges that gait variability in PD poses for state estimation by employing machine-learning models for gait-phase estimation within our exoskeleton controller. Using machine-learning-based gait-phase models deployed on a hip exoskeleton (N = 7), we performed a 2-day protocol for people with PD where the first day focused on acclimation to the device and the second focused on evaluating the device by collecting gait metrics. Using 2-min walking tests, we assessed the impact of two different types of fixed torque assistance profiles on spatiotemporal and kinematic gait metrics. We demonstrated significant improvements to hip range-of-motion (8.4%), swing time (4.7%), and peak toe clearance (12.3%) in people with PD when walking with a combined flexion and extension assistance profile as compared to walking without an exoskeleton. Although we saw trends, there were no significant differences from providing only flexion assistance given our sample size. We also demonstrated that participant-specific models reduced gait-phase estimation error by 40%, however, resulting gait metrics were not significantly altered compared to metrics when walking with the generic model. These results demonstrate that ML gait-phase-based control approaches with limited PD-specific data can improve PD gait kinematics, with enhanced accuracy associated with participant-specific data. Ultimately, these results contribute to the goal of assistive exoskeletons in everyday use for people with Parkinson’s disease.