A still from the “Supplementary Video 1” published along with the study (https://doi.org/10.1038/s41928-025-01340-2)
Brain signals in, mechanical responses out—traditionally, that’s been the limit of many brain computer interfaces (BCIs). But in a sense, this BCI doesn’t just listen; it talks back. In a study published in Nature Electronics, a research team in China reported a two-way setup that not only efficiently decodes a user’s intentions but also sends tailored feedback to shape brain activity in real time.
The technology makes use of a memristor, essentially an electronic component that can “remember” past voltage or current by changing its resistance. That capacity makes it useful for mimicking synapses in neuromorphic circuits.
The memristor-based neuromorphic decoder can thus trade bulky processors for a compact system that consumes over 1,643 times less energy than a comparable CPU-based system while learning on the fly. That results in smoother, quicker control of a drone across four degrees of freedom, and a taste of genuine co-evolution between human and hardware.
Our work is the first to introduce the concept of brain-computer co-evolution and successfully demonstrate its feasibility, marking an initial step towards mutual adaptation between biological and machine intelligence.
Piloting a drone with thoughts
In one demonstration, volunteers using the BCI were able to steer a drone through a three-dimensional obstacle course using their minds. By focusing on specific visual cues on a screen, users were able to control the drone’s movement in four degrees of freedom. That is, they were able to pilot the drone up, down, left, right, forward, backward, and rotation, all in real time. The BCI responded to both their intended movements as well as subtle error signals in their brain activity.
The paper also notes that the memristor-chip-based BCI drew offered 216× higher normalized speeds than the CPU-based system. Based on a 128k-cell memristor chip, the interface reportedly “allows the interface to achieve software-equivalent decoding performance.”
A schematic depicting the brain-memristor interactive update framework and its ability to enable a real-time brain-controlled drone flight using memristor EEG decoding. [Figure from Nature Electronics]
The 128k-cell memristor chip at the center of this interface drives a one-step decoding strategy that sidesteps traditional von Neumann bottlenecks. This one-step strategy combines preprocessing, feature extraction and pattern recognition into a single matrix operation. The result is a reduction in computational complexity and error accumulation. By analyzing steady-state visually evoked potentials (SSVEPs) and continually updating its own mapping, the system has shown a 20% jump in accuracy over static decoders, which can come in handy for real-world tasks where brain signals drift with time.
Conventional BCIs often struggle to keep pace with a fluctuating brain. But here, the memristor-based decoder and the user’s neural signals learn from each other in an ongoing loop, tightening control and minimizing energy costs. That approach could potentially find use in broader applications—from clinical rehabilitation tools to advanced human–machine collaborations—where adaptive, low-power performance is key.
The closed-loop aspect of the BCI
This closed-loop design of the BCI enables a sort of a “brain–computer co-evolution” cycle where both the decoding hardware and the user’s neural signals continuously co-adapt, which the South China Morning Post highlighted. After an initial offline training, the result is a stable, improving control interface: as the system operates, the decoder becomes more attuned to the user’s brain signals, and the user’s brain in turn learns how to produce clearer commands, leading to significantly improved communication efficiency. This adaptation is made possible through an interactive update scheme, where detected ErrPs signal incorrect decoding, allowing the system to collect new data and refine its model during use.
While the Nature Electronics paper focuses on the technical demonstration of the BCI and doesn’t detail specific real-world applications, this two-way adaptive BCI holds promise for an array of applications ranging from medicine and defense to consumer technology. The paper suggest that such systems could be generalized to a range of energy-efficient BCIs. Examples could include those using invasive recordings such as spikes electrocorticography. This, in turn, opens possibilities for applications in closed-loop neural modulation, systems that can both read brain activity and deliver stimulation to the brain in a feedback loop, potentially treating neurological disorders. In terms of rehabilitation, future systems could help patients with motor control recovery or other neurological rehabilitation needs. In addition, they could be used in brain-to-text communication, speech synthesis, motor control (controlling external devices, such as prosthetics), spinal cord stimulation for patients with spinal cord injuries as well as in locked-in patients.
More broadly, the concept of brain-computer co-evolution implies a closer merging of biological and machine intelligence. Long-term, such BCIs could reshape communication and human capabilities. Direct brain-to-brain communication or collective cognition become possibilities if extended to networks. Even at the individual level, the continuous feedback loop might enhance cognitive functions, essentially training the brain alongside AI.
