Reach for a cup and nobody experiences the calculation. You do not estimate a velocity vector, update it, and then command the fingers. You simply move toward the cup. For decades, that smoothness has hidden an awkward question from brain-computer interface (脑机接口) researchers: does the brain represent where the hand is, or only how it is moving?
A team in Beijing has an answer, and it comes with a number that the field will argue about.
The experiment behind the finding
The Institute of Automation of the Chinese Academy of Sciences (中国科学院自动化研究所) reported on 13 November that its researchers had identified what they call a three-dimensional hand-position map in the motor cortex of macaques. The work, led by Yu Shan (余山), who directs the lab for brain mapping and brain-inspired intelligence, was published in the Journal of Neuroscience.
The first author is Cao Shenghao (曹盛浩), a PhD graduate of the institute who now heads neural encoding and decoding at Zhongke Yixiang (Nanjing) Technology (中科意象(南京)科技有限公司).
Earlier motor-cortex studies often trained animals to reach toward a small set of fixed targets. That design ties position to direction: if the monkey always reaches left for one target and right for another, a neuron's apparent "position" signal may really be a direction signal in disguise. The team's fix was deceptively simple. It let the animals reach freely in three dimensions, deciding when and where to move. The same point in space could be approached from several directions, so the two variables came apart statistically.
The researchers then implanted microelectrode arrays in the left primary motor cortex (初级运动皮层, M1), which executes movement, and the dorsal premotor cortex (背侧前运动皮层, PMd), which helps plan it, and recorded individual neurons.
What the cells turned out to encode
About 36.2% of neurons in M1 and 21.3% in PMd showed clear tuning to the hand's instantaneous position. These cells do not simply fire faster the farther the hand moves. They behave like local hotspots: a given neuron becomes active when the hand enters a particular region of space. The team describes the position fields as elongated ellipsoids, and says a three-dimensional Gaussian model fitted the data better than a flat linear gradient.
The two maps are organised differently. In M1, position fields line up along the main axes of space and cluster at several scales — Cao compares it to a city with districts and buildings inside them. In PMd they are more scattered, closer to random. The team's interpretation is that M1 needs a precise coordinate framework for fine hand control, while PMd represents movement at a more abstract, planning level.
The cells also appeared to encode the intended target of a grasp, not only the hand's current location. That is a distinction with consequences: a system that can read "where the hand is going" could, in principle, leave the details of execution to an external device.
The number that matters
The most quoted result is about decoding efficiency. The team reconstructed three-dimensional hand paths from different subsets of neurons.
Using a stricter decoding subset, the picture is different: in PMd, the position-tuned cells made up just 17.7% of recorded neurons but achieved 94.1% of the decoding performance of the whole population. In M1, 32.8% of cells reached 92.9% of full-population performance. Pushed further, the analysis suggested that on average about 2.8 position-tuned cells in PMd, or 16.4 in M1, were enough to reach 90% of what the entire recorded population could do.
That result needs its warning attached. Cao told the Global Times these are statistical averages, not a claim that a working implant could use three electrodes. It describes how efficiently the brain encodes position — sparse neural coding (稀疏神经编码) — not a hardware recipe.
Conventional motor brain-computer interfaces usually decode direction and speed first, then rebuild a trajectory step by step. If position information can be read more directly, some of those intermediate steps might be skipped. Fewer channels would mean less tissue displaced, less data to transmit and potentially a longer-lived implant. Those are attractive prospects for companies trying to move beyond the small number of patients who currently receive neural devices.
Why this is still a long way from a clinic
The study was done in macaques. Human motor cortex is similar in broad outline, but nobody has yet shown that the proportion and organisation of position-tuned cells match across species. That is the first gap.
The second is time. An implant intended for a person with spinal cord injury or amyotrophic lateral sclerosis would need to work for years, possibly decades. Electrode biocompatibility, signal degradation and drift in neural signals remain unsolved problems. A beautiful decoding result over a recording session is not the same as a stable result over a decade.
The third is scope. Cao was explicit about what the finding does not support: fine manipulation, natural multitasking, or stable operation without recalibration.
For context, brain-computer interface was named a future industry in this year's Government Work Report, and a China-led reference architecture standard was published by the ISO and IEC in October. A 2025 industry report counted more than 200 Chinese companies in the field and projected a domestic market of RMB 10–14 billion (≈ US$1.4–2.0B / HK$10.7–15.2B) by 2030. None of that turns a macaque experiment into a therapy.
Key takeaways
- A Chinese Academy of Sciences team found that roughly 36.2% of M1 neurons and 21.3% of PMd neurons in macaques encode the hand's instantaneous position in three dimensions.
- A small share of position-tuned cells reproduced most of the decoding performance of the full neural population — a result about neural coding efficiency, not a proven electrode count.
- The study was published in the Journal of Neuroscience and reported on 13 November 2026, led by Yu Shan with Cao Shenghao as first author.
- The first author and the reports stress that signal stability, cross-species differences and the inability to perform fine manipulation keep this well short of clinical use.
Honest limitations
This is an animal study, and no human data are reported here. The percentages and decoding figures come from the paper as described by the institute and by the Global Times and China Daily; the paper itself was not read in full for this article. The "2.8 neurons" figure is an average the first author explicitly warned against reading as a literal hardware recipe. Market and policy figures come from a 2025 industry report and official documents and are context only. Currency conversions are approximate. The article does not claim the approach is superior to existing decoding methods, only that it changes where researchers may look.
What readers can do now
When you see a brain-computer interface story promising "thought control," ask which animal or human the data came from and over how long a session. If you follow the sector for investment or policy reasons, separate the basic-science layer — papers on coding and decoding — from the clinical and regulatory layer of approved devices; they move on different clocks. And if you are a patient or a carer, treat this result as a reason for cautious interest, not a treatment option: the hospital, not the journal, decides what is available.
