A few weeks ago we uploaded the article “Human-in-the-loop optimisation in robot-assisted gait training” to our website. The paper from the SLMC Group at the University of Edinburgh, asks a clear question. Can online optimisation really personalise robotic assistance during gait in a meaningful way for rehabilitation? The study uses Technaid’s Exo-H3 robotic exoskeleton as an experimental platform for controller personalisation. Instead of applying a single “one size fits all” controller, the authors tune parameters for each subject.
Experimental approach and control strategy
Technically, the work builds on a multi-joint impedance controller that follows an assist-as-needed principle. An evolutionary optimisation algorithm adjusts stiffness parameters based on performance metrics measured during walking. Each participant performs several sessions with the exoskeleton while the optimiser searches for an individualised setting. The outcome is a different assistance profile for every subject, which suggests a certain degree of true personalisation.
The key result, however, appears when these profiles are validated. Parameter sets converge, but functional metrics do not show clear improvements over reference configurations. This contrasts with previous studies on ankle or hip exoskeletons, where metabolic cost reductions were reported. The paper argues that human–robot coadaptation and intra-subject variability can mask the expected benefits. Optimising controller gains without modelling changes in motor strategy therefore has an obvious ceiling.
Implications for robotics and rehabilitation
From this angle, human-in-the-loop optimisation in robot-assisted gait training complements classical work on assistive control and motor learning. Other groups have shown that the choice of objective function strongly shapes outcomes. Some focus on energy cost, others on gait symmetry or muscle activation patterns. In complex systems like a multi-joint exoskeleton, optimising a single criterion may simply be insufficient for clinically relevant rehabilitation goals.
The message for the robotics and rehabilitation community is twofold. First, advanced research exoskeletons make it possible to test sophisticated personalisation strategies in controlled experimental settings. Second, tuning stiffness alone does not guarantee measurable functional gains. The next step is to combine online optimisation with explicit models of patient learning and with multi-objective cost functions closer to clinical outcomes. Only then can personalisation move beyond parameter tuning and become a real therapeutic tool
We’re eager to learn what the next steps will be for our Exo-H3. We keep moving!.