Controlling a robotic arm, a prosthetic hand, or a rehabilitation device is harder than it looks. Picking up an egg, for example, requires just the right amount of force: too little and it falls, too much and it breaks.
For people using prosthesis or patients recovering from stroke, this kind of fine control can be especially difficult because visual and tactile feedback are often reduced or absent. The less feedback is available, the harder it is to control movement accurately.
Researchers have long tried to address this challenge by including vibrations, sounds, or visual cues that stand in for what a limb would normally feel. These "augmented sensory feedback" approaches can help, but they often require additional hardware and still provide only an incomplete replacement for natural sensation.
A team led by Pierre Vassiliadis and Friedhelm Hummel at EPFL's Neuro-X Institute, with the team of Silvestro Micera and Solaiman Shokur, tested a simpler idea: instead of trying to recreate missing sensations, could they help the brain learn from success as it happens?
Constant feedback
Most training approaches tell users if they have succeeded only after a movement is complete. But a final score or success message cannot reveal which part of a complex action went wrong."
Pierre Vassiliadis, EPFL's Neuro-X Institute
The EPFL team instead designed a way to provide success information during movement. In five studies with 106 participants, including 18 chronic stroke patients, they asked participants to track a moving target for seven seconds with a cursor controlled by squeezing a force sensor or by contracting their biceps.
As participants tracked the target, its colour changed in real time according to their recent performance: green for success, red for failure. The signal adapted as participants improved, keeping the task challenging and the feedback meaningful. In control experiments, the colours changed randomly and participants were told to ignore them.
The result was striking: fewer than 20 practice trials with this simple colour feedback produced immediate improvements in motor control and these gains persisted after the feedback was removed.
Not everyone responds equally
The "color" approach actually worked best when other sources of feedback were limited. When participants could only see the cursor one third of the time, the performance benefit was roughly three times larger than when they had full visual feedback.
A similar pattern emerged in a separate experiment using a muscle-activity interface, where the benefit increased when artificial touch feedback was reduced.
Stroke patients also improved under low vision conditions, although their gains didn't persist once training stopped. The researchers suggest this may be due to the short training duration and to differences in how motor memories form after a brain injury.
Not everyone responded equally. Participants with higher reward sensitivity-a personality trait linked to the brain's reward system-showed larger improvements, both among healthy volunteers and among stroke patients. This suggests it may one day be possible to predict which patients are likely to benefit from this kind of training.
The team analyzed how information flowed between participants and the interface and found that real-time reinforcement helped to compensate for the loss of moment-to-moment motor corrections when sensory input was sparse. Rather than encouraging users to explore new strategies after making mistakes, the colour cue helped them exploit and consolidate actions that were already working.
"Because of its simplicity, the method could be added to many existing prosthetic, rehabilitation, and human-machine interface systems at little extra cost," says Vassiliadis. "By tapping into the brain's natural capacity to learn from reward, real-time reinforcement may offer a scalable way to make motor-interface training faster, simpler, and more effective."
Other contributors
- Scuola Superiore Sant'Anna
- CNRS Bordeaux
- Università Vita-Salute San Raffaele
- University Hospital of Lausanne (CHUV)
- University of Geneva
Source:
Journal reference:
Vassiliadis, P., et al. (2026). Real-time reinforcement for human-machine interface control. Neuron. DOI: 10.1016/j.neuron.2026.05.009. https://www.cell.com/neuron/fulltext/S0896-6273(26)00380-6