“Exoskeletons are already being used—but mostly in clearly defined work environments. Typically, they support repetitive movements in logistics or production and are preconfigured precisely for those tasks,” says Julian Rodemann. Soft exosuits—the lighter variant—work similarly: They are usually configured for specific tasks, such as repetitive lifting or sorting. They work less well when users need to alternate between different activities. According to Philipp Arens, a doctoral researcher at the Harvard John A. Paulson School of Engineering and Applied Sciences, this is the core challenge: “A key challenge lies in making exosuits work seamlessly across a variety of users and tasks. From a design perspective, devices can be made more lightweight and less obtrusive—but the real challenge is determining how much and at what moment in time assistance should be provided. That varies individually. This is why user feedback within the assistance loop becomes so important.”
Why Optimal Settings Are So Hard to Find
To determine which settings are optimal for which person, the researchers rely on computational support. “Real-world test sessions often tend to be lengthy,” explains Rodemann. “Participants perform different movements while my colleagues at the Biodesign Lab at Harvard University continuously record physiological data. As a result, the number of testable combinations is heavily limited.” The team therefore uses Bayesian optimization—a machine-learning technique that identifies optimal configurations efficiently by targeted sampling and progressively reducing uncertainty.
Why the AI Tests Not Only What Works Best—But Also What It Doesn’t Know Yet
In each iteration, the algorithm does not simply select the setting that appears to be the best. It must also explore unknown regions of the parameter space—even if a given configuration is temporarily less comfortable for the user. “We distinguish between exploitation, meaning the use of existing knowledge, and exploration, the targeted testing needed to close knowledge gaps,” Rodemann explains. This balance is crucial for adaptive exosuits. “Many optimization methods operate as black boxes. Users don’t know what configuration is being proposed—or why. If we can clearly and granularly explain what the system plans to do next and for what reason, trust increases. Users can also judge for themselves which regions of the parameter space do not make sense—and we avoid unnecessary tests,” says Arens.
ShapleyBO: A Method That Reveals Why the AI Makes Certain Decisions
To achieve this transparency, Rodemann and colleagues developed ShapleyBO—a method that makes optimization decisions interpretable. “Until now, humans only saw the outcome—for example, that ‘Assistance Level 7 when bending and Level 3 when lifting’ has been selected. With ShapleyBO, we show which parameters contributed to the recommendation. We also explain whether the suggestion results from optimization or from deliberate exploration of new settings. This allows users to judge whether the suggestion truly makes sense in their situation and intervene if necessary,” says Rodemann. The previously abstract balance of exploration and exploitation thus becomes visible to users.
More Studies Needed to Enable Effective Knowledge Sharing Between Humans and AI
“The algorithm knows patterns from many user datasets, while the human knows their immediate situation best. The question is how to combine both efficiently,” says Rodemann. The method is still under development and has so far been tested using simulation data from a real soft exosuit. Next, user studies will investigate how people interact with explainable optimization suggestions. “Human-in-the-loop approaches are very powerful, but they are labor-intensive and can be demanding for participants. A promising next step is to develop group-specific starting points—‘warm starts’—based on typical user profiles. This could accelerate optimization or, in some cases, even eliminate the need for it entirely,” Philipp Arens explains. Rodemann sees a fundamental advancement in this: “Our approach has the potential not only to improve the personalization of exosuits but also to strengthen trust in AI-based assistance systems overall.”
If you would like to learn more about Julian Rodemann and Philipp Arens' research, take a look at these research papers on the topic:
"Explaining Bayesian Optimization by Shapley Values Facilitates Human-AI Collaboration For Exosuit Personalization"
Authors: Julian Rodemann, Federico Croppi, Philipp Arens, Yusuf Sale, Julia Herbinger, Bernd Bischl, Eyke Hüllermeier, Thomas Augustin, Conor J. Walsh, Giuseppe Casalicchio
"Preference-based assistance optimization for lifting and lowering with a soft back exosuit"
Authors: Philipp Arens, D. Adam Quirk, Weiwei Pan, Yaniv Yacoby, Finale Doshi-Velez, and Conor J. Walsh
