Preference-Based Learning for Exoskeleton Gait Optimization (1909.12316v3)

Abstract: This paper presents a personalized gait optimization framework for lower-body exoskeletons. Rather than optimizing numerical objectives such as the mechanical cost of transport, our approach directly learns from user preferences, e.g., for comfort. Building upon work in preference-based interactive learning, we present the CoSpar algorithm. CoSpar prompts the user to give pairwise preferences between trials and suggest improvements; as exoskeleton walking is a non-intuitive behavior, users can provide preferences more easily and reliably than numerical feedback. We show that CoSpar performs competitively in simulation and demonstrate a prototype implementation of CoSpar on a lower-body exoskeleton to optimize human walking trajectory features. In the experiments, CoSpar consistently found user-preferred parameters of the exoskeleton's walking gait, which suggests that it is a promising starting point for adapting and personalizing exoskeletons (or other assistive devices) to individual users.

• The paper introduces a novel algorithm that tailors exoskeleton gait through iterative user feedback.
• It employs a mixed-initiative approach combining dueling bandits, coactive learning, and PHZD for dynamic stability.
• Simulation and human trials demonstrate that optimal gait settings are reached within roughly 20 iterations.

Preference-Based Learning for Exoskeleton Gait Optimization: A Technical Summary

The academic paper titled "Preference-Based Learning for Exoskeleton Gait Optimization" addresses a significant challenge in assistive robotics, specifically the optimization of lower-body exoskeleton gaits to enhance user comfort. This problem is approached using preference-based interactive learning, circumventing the drawbacks of conventional numerical optimization methods. The approach is particularly relevant in the context of exoskeletons, where user comfort is paramount but difficult to quantify in purely mechanical terms.

Key Concepts and Methodology

The core contribution of the paper is the development and application of a preference-based learning algorithm that leverages user feedback to tailor the walking experience provided by an exoskeleton device. By implementing a mixed-initiative approach, the algorithm called enables users to express preferences between different gaits and recommend improvements iteratively. The framework is grounded in concepts from dueling bandits and coactive learning, providing a robust theoretical foundation for preference elicitation and incorporation into the model's optimization process.

Gait Generation and Human-Machine Interface

The paper utilizes the Atalante exoskeleton, a device equipped with 12 actuated joints designed to aid individuals with lower-limb mobility impairments. The gait optimization utilizes the partial hybrid zero dynamics (PHZD) method, historically used for bipedal robots, to ensure dynamically-stable locomotion while incorporating user-comfort preferences. The optimization involves defining virtual constraints, which are dynamically adjusted based on user feedback, as opposed to relying solely on mechanical metrics like the cost of transport.

Simulation and Experimental Validation

The authors validate the algorithm in both simulation and real-world settings. Simulation experiments demonstrate the algorithm’s capacity to efficiently identify optimal gaits with minimal trials, a testament to its practicality in real-time applications where extensive testing is infeasible. Gait optimization is conducted on a simulated compass-gait biped and synthetic 2D functions to assess the algorithm's efficacy in various scenarios.

In human subject experiments, the algorithm was applied to the Atalante exoskeleton to determine optimal step lengths and other gait features. Users provided feedback on their comfort, which the algorithm used to adjust the gaits iteratively. The results confirmed that the personalized approach could discern user-preferred gaits within a limited number of trials (approximately 20), supporting the case for broader application in assistive device personalization.

Implications and Future Directions

This research highlights important implications for the design of assistive robotic systems and human-robot interaction interfaces. By embracing user preferences, rather than solely mechanical objectives, the framework represents a shift towards more human-centric design approaches in robotics. The personalization of exoskeletons can significantly enhance user satisfaction and adherence, potentially extending the benefits of such technologies to a broader population.

Future research directions include scaling the preference-based learning approach to accommodate larger feature spaces and exploring methods to generate novel gaits dynamically without a precompiled gait library. Integrating preference learning with reinforcement learning could also yield more adaptable and responsive assistive devices.

Overall, the development of preference-based gait optimization presents a valuable advance in the field of assistive robotics, aligning technological capabilities with user needs more closely and effectively.