Abstract

Soft actuators, including pneumatic artificial muscles, dielectric elastomers, and cable-driven soft fingers, exhibit highly nonlinear, history-dependent dynamics that challenge conventional PID and model-predictive controllers. Neuromorphic computing, which processes information as sparse spike trains on event-driven silicon, offers a different computational substrate: low power, high temporal resolution, and native compatibility with the asynchronous sensor signals that soft systems produce. Research groups are now coupling Intel Loihi 2 and BrainScaleS-2 neuromorphic chips directly to soft actuator control loops, demonstrating energy-efficient, adaptive motor control that scales to the small power budgets of untethered soft robots.

Why Conventional Controllers Struggle with Soft Actuators

Soft actuators derive their appeal from compliance and adaptability, properties that make them safe for human contact and robust to unmodeled collisions. However, these same properties introduce control challenges that conventional approaches handle poorly. Pneumatic muscles exhibit pressure-flow hysteresis: the force output at a given pressure depends on whether pressure is rising or falling, and this history dependence invalidates simple feedforward models. Dielectric elastomers add viscoelastic creep, where strain drifts under constant voltage over timescales of seconds to minutes.

Model-based controllers that attempt to capture this behavior require frequent online parameter updates, adding computational load that conflicts with the tight power envelopes of untethered platforms. Data-driven approaches using deep networks mitigate modeling error but carry inference costs that are similarly hard to reconcile with milliwatt budgets in centimeter-scale robots.

Neuromorphic Control Architectures

Spiking Neural Networks (SNNs) on neuromorphic hardware process information only when input spikes arrive, consuming energy proportional to activity rather than clock cycles. For soft actuator control, this maps naturally onto the event-driven output of soft sensors: capacitive strain sensors and piezoelectric films produce signals whose rate and timing encode deformation state, and SNNs can process these directly without analog-to-digital conversion overhead.

The control architecture demonstrated by the Soft Robotics Lab at EPFL in collaboration with Intel Labs pairs a pressure-based soft finger with a 64-neuron leaky integrate-and-fire (LIF) network running on Loihi 2. The SNN learns a mapping from fingertip deformation spike trains to motor command pulses through reward-modulated spike-timing-dependent plasticity (R-STDP), an online learning rule that requires no batch gradient computation. After 20 minutes of closed-loop training, the network achieves stable position control of the soft finger with average power consumption below 4 mW for the neuromorphic chip, compared to 340 mW for an equivalent ARM Cortex-M7 running a discretized nonlinear MPC at the same control rate.

Adaptation and Fault Tolerance

A practical advantage of SNN-based controllers is graceful degradation under partial hardware or sensor failure. Because information is distributed across many neurons rather than concentrated in a few critical parameters, loss of a fraction of the spike channels degrades performance gradually rather than catastrophically. This mirrors the redundancy of biological motor control and is relevant for soft robots deployed in uncontrolled environments where sensor damage is plausible.

Online adaptation through R-STDP also allows the controller to compensate for actuator aging, the gradual stiffness increase in silicone-based soft fingers as the polymer crosslinks further over months of use, without explicit model updates. This self-tuning behavior reduces maintenance requirements in long-deployment scenarios such as assistive devices and agricultural end-effectors.

Conclusion

Neuromorphic motor control is moving from a conceptual match to soft actuator properties toward demonstrated practical implementations. The combination of event-driven processing, milliwatt power consumption, and online learning through spike-timing rules addresses the three principal difficulties of controlling compliant, history-dependent actuators in untethered platforms. As neuromorphic chip ecosystems mature and soft sensor-actuator co-design practices develop, this pairing is likely to define a distinct class of low-power adaptive robots over the next several years.