Slope-Aware Variable Admittance for a Robotic Guide Dog

Slope-Aware Variable Admittance for a Robotic Guide Dog Federico Esposito PRISMA Lab - CeSMA Universit` a degli Studi di Napoli Federico II [email protected] Fabio Ruggiero PRISMA Lab - DIETI Universit` a degli Studi di Napoli Federico II [email protected] Abstract—This extended abstract investigates how terrain slope influences human preferences in physical interaction with a robotic guide dog for visually impaired individuals. A quadruped robot equipped with a rigid handle guides users through haptic feedback while varying admittance control parameters according to slope. We introduce a design combining slopeaware admittance, an energy tank for passivity, path following, and interaction force estimation. A human subjects study with blindfolded participants confirms that stiffness modulation is crucial for safety and comfort, while damping plays a minor role. These results offer design guidelines for future assistive robots. Index Terms—assistive robotics, robotic guide dog, humanrobot interaction, variable admittance I. INTRODUCTION Over 2.2 billion people worldwide are blind or visually impaired (BVI) [1]. Traditional aids such as white canes or guide dogs present limitations: canes only detect obstacles upon contact, while guide dogs are expensive (about $50,000), scarce, and limited in adaptability [2]. Robotic guide dogs are emerging as a viable alternative, offering features such as GPS navigation, AI-based perception, and long-term affordability. Robots must, however, provide safe and natural haptic interaction through a rigid handle. In the field of walking assistance [3] it was proven that users’ preferences in terms of force interaction change as the terrain slope varies. We expand such a concept as a means to improve guidance quality and safety. This work is an excerpt from [4]. II. METHODOLOGY Our framework combines a passive variable admittance filter, path following, and a momentum-based estimation of the external wrench to provide robotic guidance. A. Variable Admittance Filter The robot adapts the linear components of the stiffness and damping parameters of the admittance filter based on the estimated slope. Several variation policies are considered: intuitively we expect that the robot should stick closely to the nominal path uphill to better guide the user, and allow more compliance downhill to avoid excessive forces that may destabilize the human, but we also test a policy opposite to this intuition and a case in which the parameters are constant. This project has received funding from the ARIEL project, in the frame of the PRIN 2022 research program, grant number 2022WS29WP, funded by the European Union Next-Generation EU. B. Energy Tank for Passivity Varying admittance parameters may inject energy into the system. An energy tank monitors energy flow, storing dissipated energy and releasing it for stiffness variation if enough has been stored, or otherwise keeping the stiffness constant if the tank doesn’t have enough energy. This guarantees passivity with respect to the pair of the compliant system’s velocity and the external wrench, regardless of parameter changes. C. Path Following The reference path is defined geometrically with straight and curved segments. The robot continuously updates its reference pose based on proximity to the path, avoiding time-based trajectories that could pull the user excessively if they pause. D. Momentum-Based Estimator Interaction forces at geometric center of the robot are estimated without additional sensors by employing a momentumbased observer. This enables measurement of user input while keeping the robot hardware unchanged. III. HUMAN SUBJECTS STUDY A commercial quadruped robot (ANYmal D from ANYbotics) was equipped with a rigid handle. Outdoor trials were performed on flat terrain and an 8.5◦slope. Fifteen blindfolded participants (mean age 29.07, standard deviation 3.88) interacted with the robot. The factors considered in our experiments are (i)Path, the path the robot guides the human on, chosen between uphill, downhill and curve; (ii)Stiffness, the strategy for varying the admittance filter’s stiffness, increasing it uphill, increasing it downhill or keeping it constant; (iii)Damping, following the same options as the stiffness. To avoid having to test all the 27 possible combinations, a Taguchi L9design [5, Chapter 4.2] was employed. Participants performed the nine tests and rated Control (C), Fatigue (F), Safety (S), Aggressiveness (A), and Frustration (R). The measure (I) of how often the interaction force surpassed a set threshold was also analyzed. A. Results Table I contains the mean results across all participants for each metric and test, while Tables II-VII show the results of the ANOVA tests, performed following [5, Chapter 4.2]. 2025 I-RIM Conference October 17-19, Rome, Italy ISBN: 9788894580570 10.5281/zenodo.17629828 193 TABLE I SCORES ACROSS FACTORS FOR DIFFERENT TESTS C F S A R I Test 1 3.40 1.73 3.40 2.80 1.73 0.77 Test 2 4.00 1.40 3.60 2.73 1.53 0.78 Test 3 4.20 1.27 3.67 2.07 1.33 0.71 Test 4 3.07 1.53 3.20 2.93 1.53 0.98 Test 5 3.53 1.47 3.60 2.60 1.73 0.98 Test 6 3.53 1.80 3.53 2.60 1.60 0.97 Test 7 3.93 1.47 4.07 2.13 1.53 0.64 Test 8 4.00 1.13 4.20 2.20 1.13 0.71 Test 9 3.60 1.27 4.07 2.87 1.53 0.79 TABLE II N-WAY ANOVA ON CONTROL Source Sum Sq. d.f. Mean Sq. F p-value Path 0.4573 2.0000 0.2286 2.8405 0.2604 Stiffness 0.1788 2.0000 0.0894 1.1104 0.4738 Damping 0.2440 2.0000 0.1220 1.5153 0.3976 Error 0.1610 2.0000 0.0805 Total 1.0410 8.0000 TABLE III N-WAY ANOVA ON FATIGUE Source Sum Sq. d.f. Mean Sq. F p-value Path 0.1462 2.0000 0.0731 1.5258 0.3959 Stiffness 0.0484 2.0000 0.0242 0.5052 0.6644 Damping 0.0899 2.0000 0.0449 0.9381 0.5160 Error 0.0958 2.0000 0.0479 Total 0.3802 8.0000 TABLE IV N-WAY ANOVA ON SAFETY Source Sum Sq. d.f. Mean Sq. F p-value Path 0.7654 2.0000 0.3827 193.7500 0.0051 Stiffness 0.0365 2.0000 0.0183 9.2500 0.0976 Damping 0.1017 2.0000 0.0509 25.7500 0.0374 Error 0.0040 2.0000 0.0020 Total 0.9077 8.0000 It can be seen that Path is weakly significant in C. No factor is significant in Fand R.Path and Damping are significant in S, while Stiffness is weakly significant. Stiffness is weakly significant in A. Finally, Path is significant in Iand Stiffness is weakly significant. We can also study which levels, with respect to the significant factors, optimize our metrics. The highest control is achieved on the downhill path, while the curved path minimizes it. Safety (S) is maximized with on the uphill path with stiffness increasing uphill and constant damping, and is instead minimum on the curved path with constant stiffness and damping increasing downhill. Aggressiveness is highest with constant stiffness and lowest with stiffness increasing uphill. Finally Ipeaks on the curved path with constant stiffness and reaches its minimum on the uphill path with stiffness increasing uphill. TABLE V N-WAY ANOVA ON AGGRESSIVENESS Source Sum Sq. d.f. Mean Sq. F p-value Path 0.1462 2.0000 0.0731 0.7115 0.5843 Stiffness 0.5017 2.0000 0.2509 2.4423 0.2905 Damping 0.0247 2.0000 0.0123 0.1202 0.8927 Error 0.2054 2.0000 0.1027 Total 0.8780 8.0000 TABLE VI N-WAY ANOVA ON FRUSTRATION Source Sum Sq. d.f. Mean Sq. F p-value Path 0.0751 2.0000 0.0375 0.4343 0.6972 Stiffness 0.0040 2.0000 0.0020 0.0229 0.9777 Damping 0.0306 2.0000 0.0153 0.1771 0.8495 Error 0.1728 2.0000 0.0864 Total 0.2825 8.0000 TABLE VII N-WAY ANOVA ON INTERACTION Source Sum Sq. d.f. Mean Sq. F p-value Path 0.1237 2.0000 0.0618 30.4348 0.0318 Stiffness 0.0083 2.0000 0.0042 2.0492 0.3280 Damping 0.0012 2.0000 0.0006 0.2976 0.7706 Error 0.0041 2.0000 0.0020 Total 0.1373 8.0000 B. Discussion These findings confirm the intuition that uphill requires firmer guidance and downhill requires softer interaction. Stiffness modulation emerged as a key factor for improving user safety and minimizing the perceived aggressiveness and the required interaction, while damping played an overall minor role. In general, the framework causes little frustration or mental fatigue regardless of the policy adopted. IV. FUTURE WORK Future studies will involve visually impaired participants to validate results beyond blindfolded volunteers. Broader ranges of stiffness values will be explored, including angular admittance adaptation for curved paths, while damping will be kept constant. Finally, integration with navigation features such as obstacle avoidance and voice interaction will move robotic guide dogs closer to real-world deployment. REFERENCES [1] Blindness and vision impairment. World Health Organization. [Online]. Available: https://www.who.int/news-room/fact-sheets/detail/ blindness-and-visual-impairment. [Accessed: 18 March 2025]. [2] H. Hwang, T. Xia, I. Keita, K. Suzuki, J. Biswas, S. I. Lee, and D. Kim, “System configuration and navigation of a guide dog robot: Toward animal guide dog-level guiding work,” in 2023 IEEE Int. Conf. on Robot. and Aut., 2023, pp. 9778–9784. [3] Y. Cho, M. Lorenzini, A. Fortuna, M. Leonori, and A. Ajoudani, “A user-and slope-adaptive control framework for a walking aid robot,” IEEE Robot. and Autom. Let., vol. 9, pp. 7310–7317, 2024. [4] F. Esposito, A. Link, and F. Ruggiero, “Understanding the design of a slope-aware variable-admittance for a robot guide dog,” Submitted to the 2026 IEEE Int. Conf. on Robot. and Autom. [5] S. Park, Robust Design and Analysis for Quality Engineering. Chapman & Hall, 1996. 194