141 Picojoules per Step: When Robots Dream with Molybdenum Disulfide

> April 28, 2026, *Nature Communications*. > > A joint team from Zhejiang University, Peking University, and Renmin University of China published a paper with a mouthful of a title: *Artificial plateau neurons with in-situ spike-malleability for rhythmic quadrupedal locomotion*. But the core message is simple: > > Artificial neurons made from 2D materials directly drive a quadrupedal robot to walk. No central processor, no deep learning model, no reinforcement training. 141 picojoules per step. > > What does a picojoule mean? A single human heartbeat consumes roughly 1 joule of energy. 141 picojoules is about 1/70-billionth of that.

---

1. What It Is: From Biological "Plateau Neurons" to Silicon-Based Substitutes

1.1 Biology's Gait Secret: CPGs and Plateau Neurons

Animals walk without thinking. A cat whose brain has been removed can still walk on spinal reflexes alone. This is the Central Pattern Generator (CPG) — a neural circuit in the spinal cord of vertebrates that autonomously produces rhythmic motor signals.

One of the core neuron types in a CPG is the plateau neuron. Its defining feature: it can enter a "plateau" firing state — not a single spike, but a sustained burst of pulses. Once initiated, the burst is self-sustaining until interrupted by an inhibitory signal.

The biological significance of plateau neurons:

• Rhythm maintenance: the persistence of plateau firing stabilizes the gait cycle
• Antagonist coordination: alternating activation of extensors and flexors requires precise phase relationships
• Energy efficiency: once in the plateau state, no continuous external input is needed to sustain firing

Conventional robot control is the exact opposite: a central processor (CPU/GPU) computes each joint's angle, velocity, and torque in real time, then issues commands. High computation, high power, long latency.

1.2 Structure of the Artificial Plateau Neuron: The PG-TS Dual Component

The paper's artificial plateau neuron consists of two core components:

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

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

The PG-TS combination forms a spike oscillator:

• PG provides the plateau (the basis of sustained firing)
• TS provides the precise timing of spike emission
• Their coupling generates rhythmic spike bursts

Key innovation: in-situ spike-malleability. Rather than being tuned through external circuits, the neuron's own firing characteristics — spike amplitude, frequency, and phase — can be modulated at runtime through physical processes at the hardware level.

1.3 2D Materials: The Roles of MoS2 and Graphene

Why 2D materials?

Molybdenum disulfide (MoS2) is a flagship of the transition metal dichalcogenide (TMDC) family. A monolayer of MoS2 is only 0.65 nanometers thick, yet has excellent optoelectronic properties and a tunable electronic structure. In neuromorphic devices, MoS2 enables:

• Memristor behavior: resistance that changes with the history of electrical signals
• Optoelectronic response: optical pulses can trigger or modulate electrical signals
• Low-power operation: quantum confinement in 2D materials reduces operating voltage and current

Graphene provides:

• High conductivity: as electrode and interconnect material
• Flexibility/transparency: a foundation for future wearable or flexible robots
• Broadband response: multimodal sensing from optics to radio frequencies

The paper does not disclose the precise division of labor between MoS2 and graphene, but based on a related August 2025 *Nature Electronics* work (Wang et al.), the team used a dynamic random-access memory and inverter combination based on monolayer MoS2 films, 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 produces one rhythm oscillator. But driving a quadruped requires coordinated multi-joint control.

The paper's solution is 2(PG-TS): two PG-TS units paired.

• One unit drives the extensor
• One unit drives the flexor
• The two activate antagonistically, 180 degrees out of phase

Multiple 2(PG-TS) circuits are deployed in a distributed manner:

• One circuit per leg
• The four leg circuits coordinate through distributed encoding
• No central processor is needed for scheduling

Extended to a four-unit circuit, the system gains dynamic spike malleability and can handle multi-joint coordination in parallel.

---

2. What It's For: Hardware Instinct at 5.62 Microseconds

2.1 Unitree Go2: From Deep Learning to Hardware Instinct

The Unitree Go2 is Unitree Robotics' second-generation quadruped. Its original control pipeline:

• Sensors collect data
• Data goes to the CPU/GPU
• A deep learning model (or MPC/WBC controller) runs
• Target angles and torques are computed for each joint
• Commands are sent to motor drivers

This chain typically has millisecond-level latency (1-10 ms). The paper's scheme compresses it to 5.62 microseconds — roughly 1000x faster.

How? Skip central computation. The distributed circuits directly generate rhythmic spikes, which after Gaussian filtering drive joint-level PD (proportional-derivative) controllers. The spikes themselves are the control signal — no additional "translation" step needed.

2.2 141.37 Picojoules per Spike: Extreme Power Compression

How good is 141.37 picojoules per spike?

For comparison:

• A biological neuron spike: roughly femtojoule (fJ) to picojoule level
• Intel Loihi neuromorphic chip: ~10 pJ/spike
• IBM TrueNorth: ~26 pJ/spike
• Conventional CPU-simulated neurons: nanojoule (nJ) to microjoule (μJ) level

At 141.37 pJ, it is 1-2 orders of magnitude above biological neurons, but 3-6 orders of magnitude below conventional CPU-based approaches. Considering this is a room-temperature, cooling-free 2D-material device, the figure is highly competitive.

More importantly: this is end-to-end power — the energy of the entire chain from neuron circuit to joint controller. Not an idealized single-neuron value, but a measured value for actually driving a robot.

2.3 Stable Locomotion and Adaptive Gait Transitions

The paper demonstrates two key capabilities:

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

Adaptive gait transitions: in a real environment (not perfectly flat ground), the system adjusts its gait based on terrain feedback. This shows the hardware CPG is not fully fixed — it has some environmental adaptability.

---

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

3.1 The Truth About "Decentralization"

The paper says "without centralized computation." But that does not mean no central processor at all.

The joint-level PD controller still requires a microcontroller (MCU) to execute proportional-derivative computations. The distributed circuits generate rhythmic reference signals; actual joint control still uses PD closed loops. The paper's "decentralized" refers to decentralized rhythm generation — no central processor computing the gait cycle — but low-level control still involves distributed MCUs.

More critically: how do the four leg circuits synchronize?

A truly distributed system needs a consensus or coupling mechanism. The paper does not detail how the four leg circuits couple — physical-layer capacitive/inductive coupling? Or a shared bus? If the latter, the bus is itself an implicit "central coordinator."

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

Is 5.62 microseconds the single-neuron response latency or the end-to-end control loop latency? The paper does not make this explicit. If it is single-neuron input-to-output latency, then the full chain (sensor → ADC → neuron circuit → driver → motor) may total tens to hundreds of microseconds — still fast, but not 5.62 microseconds.

3.3 The Stability Bottleneck of 2D Materials

This is the most critical unanswered question. Known issues with 2D-material devices:

• Lifetime: MoS2 films degrade under repeated electrical stress
• Temperature sensitivity: electronic properties of 2D materials vary significantly with temperature
• Batch consistency: monolayer MoS2 growth quality is unstable
• Environmental stability: sensitivity to oxygen and water vapor

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

3.4 Completeness of Biological Replication

The artificial plateau neuron replicates the rhythmic firing of its biological counterpart. But biological CPGs have other key features:

• Adaptation: automatic adjustment of gait parameters to terrain and load
• Multistability: the same CPG network can produce multiple gaits (walk, run, jump)
• Sensory feedback integration: real-time integration of proprioceptive and external sensing
• Fatigue and learning: parameter adjustment over long-term use

Does the artificial CPG have these? Judging by "adaptive gait transitions," it has at least basic environmental adaptability. But multistability, fatigue, and learning are not addressed.

3.5 Generalization: Rhythmic Locomotion vs. Complex Terrain

Hardware CPGs excel at rhythmic locomotion (walking, trotting), but how do they handle complex terrain (stairs, obstacles, uneven ground)?

The advantage of deep learning + reinforcement learning is: learning complex-terrain strategies from large amounts of data. The advantage of hardware CPGs is: low power, low latency, high reliability. They are complementary, not replacements:

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

A future hybrid architecture may have the hardware CPG as an "instinct layer" and software AI as a "cognitive layer."

---

4. Conclusion: From Silicon Brains to Carbon Bodies

The value of this paper lies not in how fast or how steadily the robot walks, but in demonstrating a completely different control paradigm.

For the past two decades, mainstream robot control has been "software eats everything": stronger CPUs, bigger models, more data. But this creates a structural contradiction: the stronger the compute, the higher the power and the longer the latency.

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

141 picojoules per step. One seventy-billionth of a heartbeat. This energy level means future bionic robots may not need big batteries, heat sinks, or cloud computing support. Like real animals, they could operate autonomously in the wild for long periods at extremely low energy cost.

But it will take time. Engineering, mass manufacturing, and long-term stability of 2D materials — these are not problems a *Nature Communications* paper can solve. The paper is a proof of concept, a signpost, proof that "it can be done this way."

The real challenge starts now.

---

Key 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.