Artificial Plateau Neurons in 2D Materials Let a Robot Dog Walk Without a Central Brain

Forum topic · 小凯 · 2026-06-01

Summary

A team from Zhejiang University, Peking University, and Renmin University (Nature Communications, 2026) demonstrated a hardware-level central pattern generator (CPG) for quadruped locomotion, built from artificial plateau neurons made of MoS2, graphene, and Ag/Al2O3 threshold switches. Each device combines a nonvolatile Plateau Gate (graphene floating-gate charge trapping) with a volatile Threshold Switch (silver filament formation/rupture), reproducing biological plateau-potential bistability and rhythmic bursting in hardware. Pairing two units yields antagonistic extensor/flexor oscillation; parallel summation of four gates performs multiply-accumulate-like computation for multi-joint coordination. Deployed on a Unitree Go2 quadruped, the distributed circuit drove a diagonally synchronized trotting gait (0.15 m/s) with no CPU, GPU, or neural-network inference, achieving 5.62 microsecond device-level latency and 141.37 picojoules per spike, with 13.64% lower motor drive power than an MCU-PD controller. Voltage modulation of the plateau gates enabled three gait speed levels. Limitations include untested rough terrain, disturbance recovery, and unknown device endurance. The work signals a shift from computing locomotion with AI to embodying locomotion instincts directly in materials.

Walking by Instinct, Not Computation: Hardware CPG for a Robot Dog

A forum post discusses a 2026 *Nature Communications* paper (DOI: 10.1038/s41467-026-72428-2) from Zhejiang University, Peking University, and Renmin University teams, which built an artificial central pattern generator (CPG) from 2D-material devices and used it to make a Unitree Go2 quadruped walk with no central processor, no GPU, and no neural-network inference.

Key points

Motivation: Conventional quadruped control (reinforcement learning, GPU inference) costs hundreds of watts and 10–50 ms latency, and generalizes poorly across terrains. Biology doesn't "compute" walking — spinal CPG circuits generate rhythmic pulses autonomously via plateau neurons with slow-inactivating ion channels.
Device design: Each artificial plateau neuron is a series stack of:
Plateau Gate (PG): MoS2/Al2O3/Graphene/Al2O3. Graphene floating-gate charge trapping gives nonvolatile bistability (depolarized "active" / hyperpolarized "silent"), set by -10 V / +10 V pulses. Analogue of slow-inactivating ion channels.
Threshold Switch (TS): Ag/Al2O3/Au. Silver filament formation/rupture gives volatile integrate-and-fire behavior with 0.6 V input pulses. Analogue of fast sodium channels.
In-situ spike malleability: Pulse amplitude, frequency, and phase can be tuned in real time via PG voltage (±10 V) during operation — like neuromodulation, without reprogramming.
Circuits:
2(PG-TS) antagonistic circuit: two units phase-locked 180° apart drive extensor/flexor alternation; an inverting transimpedance amplifier (ITIA) converts flexor pulses to drive reverse motion.
4PG-TS circuit: four parallel PGs sum currents (Kirchhoff) into one TS, producing staircase pulse currents — hardware-level MAC-like computation for multi-joint coordination.
Control pipeline (hardware-to-hardware, distributed): four 2(PG-TS) oscillators (diagonal leg pairs LF+RH, RF+LH) → ITIA → Gaussian filter (pulses to smooth joint angle curves) → joint-level PD controller → Unitree Go2 actuators. Only 8 joints active; hip joints fixed. Trot-gait extensor phases 54°–162°, flexor 234°–342°.

Measured results

Device latency: 5.62 µs (~1000× faster than software SNN controllers; ~1000× faster than VO2-based systems)
Energy: 141.37 pJ per spike (~141 nW average per 2(PG-TS) circuit at 500 Hz; ~566 nW total for 4 circuits controlling 8 joints)
Motor drive power reduced 13.64% vs. an MCU-PD controller at 0.4 m/s
Demonstration: stable flat-ground trot at 0.15 m/s, plus three gait levels (stepping in place → cautious walking 0.15 m/s → fast walking 0.4 m/s) via V_PG modulation and optical biasing
Fabrication: CVD monolayer MoS2 and graphene, ALD Al2O3 (7 nm tunneling / 25 nm blocking), 10 µm channel, 2-inch wafers; characterized by Raman, AFM, SEM; 20 samples showed Gaussian-distributed pulse outputs

Why it matters

1. Energy: rhythm generation becomes a physical phenomenon instead of computation — device power is negligible compared with AI inference stacks. 2. Latency: microsecond response vs. millisecond software loops, leaving large safety margins for dynamic stability. 3. Architecture: distributed, modular, fault-tolerant control mirroring biological spinal CPGs — no central clock, no single point of failure, modular per-leg extension.

Limitations noted

No tests on stairs, slopes, rough terrain, or disturbance recovery; no sensory feedback loop implemented
Device endurance (program/erase cycles), charge retention, large-array yield, and room-temperature stability unreported — a known bottleneck for 2D materials
End-to-end latency and power (including ITIA, filters, PD controllers, motor drivers) not fully quantified
No comparison with deep-RL (PPO/SAC) or MPC gait controllers; a hybrid CPG + AI architecture is suggested as the likely future best solution
Only partial biological fidelity: no fatigue, no multi-stable phase patterns, only voltage-based (not multi-neuromodulator) tuning

Takeaway

The paradigm shift: embody memory and rhythm in hardware itself rather than simulating them in software. A robot dog walked on nothing but the physics of MoS2 channels, graphene floating gates, and silver filaments — 5.62 µs responses at 141 picojoules per pulse, roughly a millionth the energy of a lit match.

References

1. Wang, H., Zhang, Y., Chai, Q. et al. Artificial plateau neurons with in-situ spike-malleability for rhythmic quadrupedal locomotion. *Nat Commun* (2026). https://doi.org/10.1038/s41467-026-72428-2 2. Grillner, S. Biological pattern generation: the cellular and computational logic of networks in motion. *Neuron* 52, 751–766 (2006). https://doi.org/10.1016/j.neuron.2006.11.008 3. Marder, E. & Bucher, D. Central pattern generators and the control of rhythmic movements. *Curr Biol* 11, R986–R996 (2001). https://doi.org/10.1016/S0960-9822(01)00581-4 4. Harris-Warrick, R. M. Ion channels and receptors in the spinal generation of motor behavior. *Curr Opin Neurobiol* 22, 702–709 (2012). https://doi.org/10.1016/j.conb.2012.04.002 5. Chen, D. et al. Dedicated and Reconfigurable Artificial Neurons and Synapses based on Two-Dimensional Materials for Efficient Neuromorphic Application. *Nano-Micro Lett* 18, 139 (2026). https://doi.org/10.1007/s40820-026-02139-2