Researchers at the University of Ljubljana and the Jožef Stefan Institute tested how the stiffness linking two KUKA LBR iiwa arms affected the process of teaching the robot to move a box. Fifteen volunteers, including 11 men and four women with an average age of 23.8 years, physically guided both robot end-effectors through repeated demonstrations. Robot experience ranged from none among some participants to master’s-level study among others. The researchers used a randomized within-subject design, so each person experienced four controller configurations, while the order of those configurations was assigned randomly.

The four settings ranged from fully independent arms to strong bimanual coupling. In the OFF condition, relative stiffness was zero, allowing the two arms to move independently. LOW used translational relative stiffness of 15 newtons per meter and orientational stiffness of 4 newton-meters per radian. MID increased those values to 70 N/m and 10 N m/rad, while HIGH used 200 N/m and 25 N m/rad. Absolute impedance was kept low so that the controller primarily governed the relationship between the arms rather than their position in the workspace. Damping was tuned separately for relative and absolute motion, and a nullspace task kept each seven-axis arm away from its joint limits.

Participants grasped opposite sides of a rigid 50.5-centimeter-long box, crossed a six-centimeter-high obstacle and placed the box at one of two marked targets. The controller maintained a fixed relative pose between the arms throughout the demonstration, including the initial grasp and final placement. Its reference was deliberately 10 centimeters narrower than the box, creating a baseline compression that made the arms continuously resist separation once both end-effectors were in contact. This design meant that stronger coupling would mechanically increase the load, so the investigators also separated the controller’s commanded wrench from the residual difference between measured and commanded forces.

The clearest objective benefit appeared in task time. Mean transport time was 22.9 seconds with independent arms, 19.4 seconds at LOW, 20.0 seconds at MID and 17.8 seconds at HIGH. A Friedman test identified a significant overall effect of controller setting, with a moderate effect size. Bonferroni-corrected post-hoc comparisons showed a significant difference between OFF and HIGH, corresponding to an approximately 22.3% reduction in mean cycle time and a large rank-biserial effect. The intermediate conditions did not produce a strictly monotonic time pattern, which means the results support a clear advantage for HIGH over OFF rather than demonstrating that every incremental increase in stiffness improves performance.

Final placement accuracy did not differ significantly among the four settings. Mean planar positioning errors were 16.9 millimeters for OFF, 16.8 millimeters for LOW, 16.1 millimeters for MID and 16.8 millimeters for HIGH. A repeated-measures analysis of variance found no significant condition effect, and none of the pairwise comparisons met the statistical-significance threshold. The investigators suggested that participants deliberately slowed their movements before release, allowing them to compensate for coordination differences during the fine-positioning phase. The null result therefore should not be read as proof that coupling cannot affect precision; the task design may simply have masked such an effect.

Interaction forces moved in the opposite direction from perceived effort. The sum of force magnitude across the two end-effectors increased monotonically as relative stiffness rose, with significant differences between every pair of settings in data from 14 participants whose force recordings were complete. The commanded controller wrench accounted for a growing share of measured variation, with median R-squared values of 0.01, 0.09 and 0.37 for LOW, MID and HIGH. Yet the residual force also grew, with medians of 1.3, 2.1, 3.6 and 6.0 newtons from OFF through HIGH. Because that residual includes operator actions as well as unmodeled actuation effects, it cannot be assigned solely to the human. Even so, its continued increase shows that the controller’s designed-in pretension does not fully explain the force pattern.

Participants nevertheless reported less workload under stronger coupling. Raw NASA-TLX scores showed lower perceived demands with higher stiffness, with significant differences between OFF and LOW, OFF with MID, and OFF with HIGH. Most subscales followed the same general improvement, although differences among the higher settings were not significant. The apparent contradiction between greater measured force and less perceived effort is consistent with a redistribution of control: at LOW, operators had to maintain fine relative coordination themselves, whereas stronger coupling let them concentrate on moving the assembly toward each target. Some participants nevertheless preferred MID over HIGH, indicating that the practical benefit of additional stiffness may plateau, reverse or vary between individuals. The study tested only a limited force range, so this pattern should not be extrapolated to substantially stronger coupling.

The experiment also had important boundaries. It examined demonstration performance rather than the autonomous behavior a robot might later learn, so it did not compare the training value of trajectories collected under each setting. There was no dedicated practice period, and the small, demographically homogeneous sample of 15 meant that intra-session learning and fatigue may not have been fully balanced despite randomization. Camera placement introduced noise into placement measurements, while fixed controller gains could not accommodate individual preferences. The findings also apply directly only to a symmetric task with a rigid object whose effective grip-point relationship remained constant. Tasks involving changing relative orientation, deformable material or different roles for the two arms could work against a strong fixed constraint and require separate testing.

For the tested pick-and-place operation, the results identify controller coupling as more than a low-level engineering parameter: it shaped how quickly people could teach a bimanual robot, how much force the interaction produced and how demanding the process felt. HIGH coupling offered the strongest contrast with fully independent operation, reducing mean task time and perceived workload without a statistically detectable change in final placement error. The results do not show that all coordination is better than independence, nor that stronger is always better. They instead support tailoring relative stiffness to the task and the demonstrator, with future work needed to test adaptive gains, broader manipulation tasks, larger and more diverse groups, and the quality of robot policies learned from the resulting demonstrations.

Subject of Research: Effects of relative stiffness on bimanual kinesthetic robot teaching

Article Title: Effectiveness of bimanual kinesthetic teaching through different controller settings

Article References: Avsec, O., Ude, A., Mihelj, M., & Gams, A. (2026). Effectiveness of bimanual kinesthetic teaching through different controller settings. Frontiers in Neurorobotics, 20, Article 1960197. https://doi.org/10.3389/fnbot.2026.1960197

Image Credits: AI Generated

DOI: 10.3389/fnbot.2026.1960197

Keywords: bimanual robotics, kinesthetic teaching, impedance control, human-robot interaction, robot learning by demonstration, dual-arm manipulation, collaborative robots, task performance, perceived workload, Effectiveness, bimanual, kinesthetic

The motivation is practical and pressing. Nursing staff spend significant time and physical effort making beds, and the repetitive bending, lifting, and reaching involved contributes to musculoskeletal strain. Automated bed-making could reduce that physical burden while also lowering infection risk, since consistent, standardized bed preparation matters in clinical environments. Surveys cited in the study suggest that nurses and patients are generally willing to work alongside service robots in healthcare settings, and recent studies indicate that nursing robots can reduce workload in general adult wards. But translating that demand into a working machine requires solving a problem that has resisted a decade of research: reliably manipulating cloth, the archetypal deformable object, in an unstructured real-world environment.

The core of the new framework is a formal twelve-step task decomposition of bed-making. Rather than treating the job as one monolithic skill, the researchers broke it into a reproducible sequence of task-level stages, bridging high-level semantic reasoning and low-level manipulation control. This structured representation matters because long-horizon tasks compound errors: a small mistake in an early stage, such as a poor grasp of a sheet corner, propagates through every subsequent action. By formalizing the task as a discrete sequence, the team created a framework in which each step can be monitored, evaluated, and recovered from independently, and in which performance can be measured stage by stage rather than only as an overall pass-or-fail outcome.

Perception is driven by semantic keypoints, a technique that has become a powerful tool for handling deformable objects. Instead of asking a vision system to label every pixel of a crumpled sheet, the robot uses vision-language models to localize semantically meaningful points: the corner of a sheet, the edge of a mattress, the midpoint of a fold. The approach builds on open-vocabulary detection systems such as Grounding DINO and YOLO-World, and on the visual grounding tradition established by CLIP, allowing the robot to identify task-relevant points from natural language descriptions without needing task-specific retraining. Earlier work on general-purpose clothes manipulation with semantic keypoints demonstrated the promise of this representation, and the new study extends it to the more demanding setting of dual-arm bed-making, where two manipulators must act on the same deformable object in a coordinated way.

Once keypoints are localized, the system plans coordinated dual-arm motion for fabric handling. Two arms are not merely a convenience here; they are essential. Spreading a sheet across a mattress, aligning its edges, and folding it flat all require simultaneous control of two distant points on the same cloth, with continuous tension maintained between them. The manipulation planner generates motions that respect the coupled geometry of the fabric, drawing on a research lineage that includes model-free visual servoing for deformation control, dynamic movement primitives for deformable manipulation, and learned pick-point detection for robotic bed-making. The team also employed hand-eye calibration using 3D-to-3D correspondences to align what the cameras see with where the arms actually move, a deceptively mundane step that is critical for millimeter-level grasp accuracy.

Perhaps the most consequential engineering contribution is the closed-loop recovery mechanism. In open-loop systems, a robot executes a planned sequence and hopes for the best; any failure, such as grasping two layers of fabric instead of one, silently corrupts everything that follows. The new system instead monitors each step, detects unsuccessful grasps, and updates its perception, planning, and execution accordingly. If a grasp fails, the robot re-perceives the scene, re-plans, and retries, treating uncertainty as an expected condition rather than an exception. This probabilistic stance toward real-world execution, long a theme in robotics, is what lifts the system’s average step-wise success rate to 91.9 percent across the twelve stages, a figure that would be unattainable with naive open-loop execution.

The experimental evaluation on a real robotic system reveals a striking asymmetry in difficulty. Rigid manipulation tasks, such as moving objects with well-defined geometry, achieved near-perfect reliability. Deformable operations, particularly sheet folding, remained the primary bottleneck, hampered by perception and grasp uncertainty. The most troublesome failure mode identified by the researchers occurs in the early stages, when the robot must distinguish individual fabric layers. Grabbing two layers of a sheet when only one is intended may look superficially successful, but the error cascades through the task, distorting subsequent folds and alignments in ways that later recovery steps cannot fully undo. This error propagation across task stages is precisely why the overall task success rate lags behind the per-step figure, and it encapsulates the central challenge of long-horizon deformable manipulation.

The study situates itself within a rapidly evolving research landscape. Deep-learned pick-point detection for robot bed-making dates back to work presented at the International Symposium on Robotics Research in 2019, and simulation-to-real reinforcement learning for deformable objects has been explored since at least 2018. More recently, large language models have been used for symbolic planning in long-horizon deformable assembly, and open-source vision-language-action models such as OpenVLA have begun to unify perception and control. Tactile sensing is emerging as a complementary channel: researchers have shown that robots can singulate layers of cloth using tactile feedback, and transformer-based multimodal frameworks are being developed for material perception. The Taiwanese team’s contribution is to integrate these threads, vision-language keypoint localization, dual-arm planning, and closed-loop recovery, into a single working system tested end-to-end on hardware, rather than in simulation alone.

The authors are candid about the limitations and the path forward. Future work will integrate tactile and visual feedback to improve layer-state estimation, giving the robot a direct physical sense of how many fabric layers it holds, and to enable adaptive recovery strategies that respond to the actual state of the cloth rather than only its appearance. This is a sensible direction, since vision alone struggles with the self-occlusion and ambiguity that make layered fabric so treacherous. The research was supported by Taiwan’s National Science and Technology Council and the Industrial Technology Research Institute, reflecting an institutional commitment to translating assistive robotics from laboratory demonstrations into clinically useful systems.

For the field of embodied AI, the 91.9 percent step-wise success rate is both an achievement and a diagnostic. It shows that structured task decomposition, semantic perception, and failure-aware control can carry a robot through a genuinely long-horizon household task on real hardware. It also shows, with unusual clarity, exactly where the frontier lies: not in rigid grasping, which is essentially solved at this scale, but in the entangled physics of cloth, where seeing is not always knowing. If the next generation of these systems can feel the difference between one layer and two, the humble act of making a bed may become one of the first full-scale deformable manipulation tasks that robots perform reliably alongside nurses, and a template for the laundry-folding, dressing-assisting, and linen-handling robots that would follow.

Subject of Research: Autonomous dual-arm robotic bed-making using semantic keypoint perception and closed-loop recovery for deformable object manipulation

Article Title: Dual-arm manipulation driven by semantic keypoints for autonomous bed-making

Article References: Shih, C.-H., Cheng, P.-H., & Chuang, Y.-Y. (2026). Dual-arm manipulation driven by semantic keypoints for autonomous bed-making. International Journal of Intelligent Robotics and Applications. https://doi.org/10.1007/s41315-026-00597-w

Image Credits: AI Generated

DOI: 10.1007/s41315-026-00597-w

Keywords: assistive robotics, deformable object manipulation, semantic keypoints, dual-arm manipulation, vision-language models, closed-loop recovery, long-horizon planning, hospital automation, robotic perception, fabric handling, embodied AI, nursing robots