--- title: "No surgery, just a headset: a Chinese lab flew a drone four ways with thought alone" date: 2026-10-15 category: Brain–Computer Interface site: NeuroAI canonical: https://neuroai.site/a/na-brain-dual-loop-drone language: en --- # No surgery, just a headset: a Chinese lab flew a drone four ways with thought alone > A Tsinghua–Tianjin team built a non-invasive "dual-loop" brain-computer interface (脑机接口) on memristor hardware that steered a drone through four degrees of freedom and got better over six hours. Most non-invasive brain-computer interfaces (脑机接口) can manage a yes-or-no, a left-or-right. A joint Tsinghua–Tianjin team just pushed that to a drone doing four things at once — and, oddly, the system improved the longer it ran. On 17 February 2025, Tianjin University's Brain-Computer Haihe Laboratory (脑机海河实验室) and Tsinghua University's School of Integrated Circuits reported a 双环路 (dual-loop) non-invasive evolving BCI built on memristor neuromorphic devices. The work was published in *Nature Electronics*. The demo: a person wearing a non-invasive cap steered a drone through **four degrees of freedom** — up-down, left-right, forward-back, and rotation. ## The problem with "non-invasive" Non-invasive BCIs read brain activity through the skull, usually with EEG electrodes on the scalp. That avoids surgery, but the signal is weak, noisy, and changes from minute to minute as the user shifts attention, sweats, or simply tires. The classic failure mode is drift. A decoder trained at the start of a session performs worse an hour later, because the brain signal it learned has wandered. Researchers have patched this by re-calibrating, but recalibration is annoying and breaks the sense of direct control. The Tsinghua–Tianjin team attacked the cause. They argued that the signal's non-stationarity — its tendency to change — is not just background noise. It is partly the brain *learning the task* in real time, guided by the闭环 (closed loop) of the interface itself. ## The dual loop The fix is the name of the paper: a 双环路 (dual-loop) framework where two learning processes run together. - The **"machine learning" loop**: a memristor-based decoder updates its parameters to track the drifting brain signal. - The **"brain learning" loop**: the user's task-related brain features evolve positively under the guidance of a decision-feedback cycle — the person gets better at producing the signal the machine wants. The two reinforce each other. Early on, the decoder does most of the adapting. Over time, the brain's contribution grows. The team reported that across a continuous six-hour interaction, performance did not degrade — it rose by about **20%** in accuracy. ## Why memristors matter here The decoder was not running on an ordinary processor. It was deployed on a **128kb-scale memristor neuromorphic device** (忆阻器神经形态器件). Memristors are components whose resistance can be set and remembered, mimicking how synapses strengthen and weaken. Used this way, they do brain-like computation in hardware. The payoff is brutal efficiency. Against a conventional all-digital scheme, the memristor approach: - Ran decoding **more than two orders of magnitude faster** — over 100 times. - Cut energy use by **three orders of magnitude** — to roughly a thousandth. - Collapsed the brain signal's multi-step computation into a single step. For a BCI, speed and power are not trivia. They determine whether a device can be worn, run all day, and respond before the user's attention drifts. ## From two degrees to four The team made a specific, useful comparison. Traditional non-invasive BCIs, they said, typically manage only simple **two-degree-of-freedom** flight — think up-down or left-right. Their dual-loop system handled four degrees at once, enough for a drone to move through space and orient itself. That is the kind of step that moves a BCI from a lab curiosity toward something usable: not just "go" or "stop," but continuous, multi-axis control. ## Where this sits in the field Non-invasive BCIs are the slow, safe lane of brain-computer interface (脑机接口) work. No operating room, no implant, no infection risk — which is exactly why they may reach ordinary users first, even if their bandwidth is lower than invasive rivals. According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, non-invasive interfaces will likely reach patients and consumers years before invasive ones, simply because no surgeon is required to start. Xu Minpeng (许敏鹏), a professor at the Haihe Lab, said the team plans to push the system into more portable and wearable BCI devices, serving both consumer and medical human-machine interaction. ## Honest limitations - **Demonstration-scale, not product-scale.** The result is a controlled lab demo of drone control; it is not a shipped device or a certified medical tool. - **Healthy users, short tasks.** Reports describe able-bodied participants in a defined task, not patients with motor impairment using it day to day. - **Efficiency claims are hardware-specific.** The 100x speed and 1000x energy figures come from comparing to a particular conventional digital baseline; real-world wearables involve many other costs. - **No long-term or at-home data.** The six-hour test is the longest window reported; durability over weeks or months is unknown. ## What readers can do now - **Watch the wearable path.** Non-invasive BCIs need no surgery, so they are the most likely to appear in clinics and homes first — track mentions of portable or headset form factors. - **Look for the energy story.** Memristor and neuromorphic hardware is the reason these systems can shrink; follow chip news, not just brain news. - **Demand the six-week test.** A 20% gain over six hours is promising; the real proof is whether it holds over weeks without recalibration. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-brain-dual-loop-drone Free to quote with attribution and a link to the original.