141 Picojoules Per Step: MoS2 Artificial Neurons Drive a Quadruped Robot Without a CPU

Forum topic · 小凯 · 2026-06-01

Summary

A joint team from Zhejiang University, Peking University, and Renmin University of China reported in Nature Communications (April 2026) an artificial plateau neuron built from two-dimensional materials that directly drives rhythmic locomotion on a Unitree Go2 quadruped robot — without a central processor, deep learning models, or reinforcement training. Each artificial neuron circuit, built from a plateau gate (PG) paired with a threshold switch (TS) on monolayer MoS2 with graphene interconnects, consumes just 141.37 picojoules per pulse, several orders of magnitude below conventional CPU-based control. The distributed 2(PG-TS) circuits compress the control latency to about 5.62 microseconds, roughly 1000x faster than millisecond-level deep-learning pipelines, and demonstrate stable trotting plus adaptive gait transitions on real terrain. This analysis also notes open questions: joint-level PD controllers still need microcontrollers, inter-leg synchronization mechanisms are unspecified, and MoS2 devices face known engineering challenges around lifetime, temperature sensitivity, and batch consistency. The work points toward a 'physical AI' paradigm where material physics itself performs computation, potentially enabling extremely low-power autonomous robots.

141 Picojoules Per Step: When Robots Dream in Molybdenum Disulfide

> Source: *Nature Communications*, April 28, 2026. Joint team from Zhejiang University, Peking University, and Renmin University of China. > > The paper's title is a mouthful: *Artificial plateau neurons with in-situ spike-malleability for rhythmic quadrupedal locomotion*. The core message is simple: > > Artificial neurons made from 2D materials directly drive a quadruped robot to walk. No central processor, no deep learning model, no reinforcement training. 141 picojoules per step. > > For scale: one human heartbeat costs roughly 1 joule. 141 picojoules is about one seventy-billionth of that.

1. What It Is: From Biological "Plateau Neurons" to Silicon-Free Replacements

1.1 The Biological Secret of Gait: CPG and Plateau Neurons

Animals walk without thinking. Even a cat whose brain is removed can walk via its spinal cord. This is the central pattern generator (CPG) — neural circuits in the vertebrate spinal cord that autonomously produce rhythmic motor signals.

A core neuron type in CPGs is the plateau neuron. Its defining trait: it can enter a "plateau" firing state — not single spikes, but self-sustaining bursts that persist until interrupted by inhibitory signals.

Biological significance of plateau neurons:

Rhythm maintenance: plateau firing sustains stable gait cycles
Antagonist coordination: alternating extensor/flexor activation requires precise phase relationships
Energy efficiency: once in the plateau state, no continuous external input is needed

Traditional robot control is the opposite: a CPU/GPU computes every joint angle, velocity, and torque in real time — computationally heavy, power-hungry, high-latency.

1.2 The Artificial Plateau Neuron: A PG-TS Two-Component Design

The proposed artificial plateau neuron has two core components:

PG (Plateau Gate): a bistable switch. Once triggered, it enters a self-sustaining "on" state producing a sustained membrane-potential plateau.

TS (Threshold Switch): a transient switch. When membrane potential exceeds a threshold it fires a spike, then rapidly resets.

The PG-TS pair forms a spike oscillator: PG provides the plateau basis for sustained firing, TS provides spike timing, and their coupling produces rhythmic bursts.

Key innovation: in-situ spike-malleability — the neuron's own firing properties (amplitude, frequency, phase) can be modulated at runtime through hardware-level physical processes, not external circuits.

1.3 The 2D Materials: Roles of MoS2 and Graphene

Molybdenum disulfide (MoS2), a representative transition metal dichalcogenide (TMDC), is only 0.65 nm thick as a monolayer yet offers:

Memristive behavior: resistance that changes with electrical history
Optoelectronic response: light pulses can trigger or modulate electrical signals
Low-power operation: quantum confinement lowers operating voltage and current

Graphene provides:

High conductivity as electrodes and interconnects
Flexibility/transparency as a basis for future wearable or soft robots
Broadband response from optics to RF

The paper does not detail the exact division of labor, but a related August 2025 *Nature Electronics* work (Wang et al.) indicates the team used monolayer MoS2-film-based dynamic random-access memory plus inverter combinations, modulating the voltage in the DRAM capacitor (i.e., the neuron's membrane potential) to emulate intrinsic plasticity.

1.4 2(PG-TS): From Single Neurons to Distributed Circuits

A single PG-TS pair is one rhythm oscillator; driving a quadruped requires multi-joint coordination. The paper's scheme is 2(PG-TS): two PG-TS units paired — one drives the extensor, one the flexor, antagonistically activated with a 180° phase difference.

Multiple 2(PG-TS) circuits are distributed — one per leg, coordinated by distributed encoding, with no central processor scheduling. Scaling to a four-unit circuit adds dynamic spike malleability for parallel multi-joint coordination.

2. What It's Good For: 5.62-Microsecond Hardware Instinct

2.1 Unitree Go2: From Deep Learning to Hardware Instinct

The Unitree Go2 (Unitree Robotics' second-gen quadruped) normally follows: sensors → CPU/GPU running deep learning (or MPC/WBC controllers) → per-joint target angles/torques → motor drivers. This chain typically runs at millisecond latency (1–10 ms). The paper compresses it to 5.62 microseconds — roughly 1000x faster.

How? Skip central computation. Distributed circuits directly generate rhythmic spikes; after Gaussian filtering, they drive joint-level PD (proportional-derivative) controllers. The spikes themselves are the control signal — no extra "translation."

2.2 141.37 Picojoules Per Spike: Extreme Power Compression

For reference:

Biological neuron per spike: ~1–10 fJ to pJ range
Intel Loihi neuromorphic chip: ~10 pJ/spike
IBM TrueNorth: ~26 pJ/spike
Traditional CPU neuron simulation: nJ to μJ range

141.37 pJ is 1–2 orders of magnitude above biological neurons, but 3–6 orders of magnitude below conventional CPU approaches — at room temperature, no cooling, on a 2D-material device.

Crucially, this is end-to-end power: from neuron circuit to joint controller, the measured cost of actually driving a robot, not an idealized single-neuron figure.

2.3 Stable Walking and Adaptive Gait Transitions

Two demonstrated capabilities:

Stable on-ground locomotion: energy-efficient trotting on flat ground — note "energy-efficient": conventional control can also walk the Go2, but at far higher power than this hardware-CPG scheme.

Adaptive gait transitions: on real (non-ideally-flat) terrain, the system adjusts gait from terrain feedback — the hardware CPG is not fully fixed and has some environmental adaptability.

3. How to Use It: The Boundaries of Decentralized Control

3.1 The Truth About "Decentralization"

"Without centralized computation" doesn't mean no processor at all. Joint-level PD controllers still need microcontrollers (MCUs). The distributed circuits produce rhythmic spike reference signals; actual joint control still requires PD closed loops. "Decentralized" here means decentralized rhythm generation — no central CPU computing gait cycles, though distributed MCUs remain at the bottom layer.

More critically: how do the four legs' circuits synchronize? True distributed systems need consensus or coupling. The paper doesn't detail this — is it physical-layer capacitive/inductive coupling, or a shared bus? If the latter, the bus itself is an implicit "central coordinator."

3.2 End-to-End Latency vs Single-Neuron Latency

Is 5.62 μs a single-neuron response delay or the full end-to-end loop? The paper is unclear. If it's input-to-output of one neuron, the complete chain (sensor → ADC → neuron circuit → driver → motor) might total tens to hundreds of microseconds — still fast, but not 5.62 μs.

3.3 The 2D-Material Stability Bottleneck

The key unanswered question. Known issues with 2D-material devices:

Lifetime: MoS2 films degrade under repeated electrical stress
Temperature sensitivity: electronic properties vary strongly with temperature
Batch consistency: monolayer MoS2 growth quality is unstable
Environmental stability: sensitivity to oxygen and water vapor

The paper does not discuss these. Between a *Nature Communications* prototype and an industrial product typically lies 5–10 years of engineering.

3.4 Completeness of Biological Fidelity

The artificial plateau neuron reproduces rhythmic firing, but biological CPGs also have:

Adaptation: automatic gait parameter tuning for terrain/load
Multistability: one CPG network producing multiple gaits (walk, run, jump)
Sensory feedback integration: real-time proprioceptive and exteroceptive fusion
Fatigue and learning: long-term parameter adjustment

From "adaptive gait transitions," the artificial CPG at least has basic environmental adaptability, but multistability, fatigue, and learning are not addressed.

3.5 Generalization: Rhythmic Motion vs Complex Terrain

Hardware CPGs excel at rhythmic motion (walking, trotting), but what about stairs, obstacles, uneven ground? Deep learning + reinforcement learning learns complex-terrain strategies from large datasets; hardware CPGs win on low power, low latency, and reliability. They are complementary, not substitutes:

Simple terrain: hardware CPG takes over, saving energy
Complex terrain: switch to software control (deep learning), trading power for capability

A future hybrid architecture could use the hardware CPG as the "instinct layer" and software AI as the "cognition layer."

4. Closing: From Silicon Brains to Carbon Bodies

This paper's value is not that it makes robots faster or more stable, but that it demonstrates a fundamentally different control paradigm.

For two decades, mainstream robot control has been "software eats everything": stronger CPUs, bigger models, more data. That creates a structural contradiction: more compute means more power and more latency.

This work returns to an older paradigm: let physics itself do the computation. MoS2's physical properties (memristance, optoelectronic response, threshold switching) directly encode neuronal dynamics — no external program simulating them. This is a model of Physical AI: not emulating a carbon-based brain on silicon chips, but implementing similar functions directly through the physical processes of new materials.

141 picojoules per step — one seventy-billionth of a heartbeat. That energy scale means future biorobots may not need big batteries, cooling, or cloud compute. They could run autonomously in the wild at extremely low energy, like real animals.

But it takes time. Engineering, mass manufacturing, long-term stability of 2D materials — these are not problems a *Nature Communications* paper solves. The paper is a proof of concept, a signpost, a proof that "you can walk this way."

The real challenge starts now.

Core References

1. Wang et al. (2026). Artificial plateau neurons with in-situ spike-malleability for rhythmic quadrupedal locomotion. *Nature Communications*, s41467-026-72428-2. 2. Wang et al. (2025). A bioinspired artificial neuron with intrinsic plasticity based on monolayer MoS2. *Nature Electronics*, DOI: 10.1038/s41928-025-01433-y. 3. Yan et al. (2021). Reconfigurable stochastic neurons based on tin oxide/MoS2 hetero-memristors for simulated annealing and Boltzmann machines. *Nature Communications*, 12, 5455. 4. Gu & Dao (2024). Mamba: Linear-Time Sequence Modeling with Selective State Spaces. *ICML*. 5. Merolla et al. (2014). A million spiking-neuron integrated circuit with a scalable communication network and interface. *Science*, 345(6197), 668-673.