Introduction
Although considerable progress has been made, robotic motor control still lacks the adaptability of biological motor systems in dynamic environments [1]. Replicating the flexibility of biological systems remains a challenge in neurorobotics. The cerebellum is crucial for motor control [2], suggesting it can inform the design of adaptive robotic controllers. Since Marr’s seminal theory [3], most cerebellar models focus on parallel fibre (PF)–Purkinje cell (PC) plasticity. However, other forms of cerebellar plasticity exist [4] and remain underexplored in real-time robotic implementations. We investigated the functional contribution of synaptic plasticity in the deep cerebellar nuclei (DCN) in real-time cerebellum-based robotic control.
Methods
We extended a validated cerebellar spiking network model for robotic control [5]. The model comprises 61,200 leaky integrate-and-fire neurons: 60,000 granule cells, 600 PCs and 600 DCN neurons. Mossy fibres (MFs) encode desired and actual kinematic signals, while climbing fibres convey error signals. The controller operates in real time with a compliant Baxter robot via intermediate modules converting analogue signals into spikes and vice versa. The robot acquires the target trajectory through activity-dependent plasticity at PF–PC synapses. As experimental studies suggest interactions between PF–PC learning and plasticity in the deep cerebellar nuclei, we implemented synaptic plasticity at MF–DCN and PC–DCN synapses.
Results
We compared the mean absolute error (MAE) of different target trajectories across several network configurations: a homosynaptic learning rule at MF–DCN synapses, a heterosynaptic learning rule at MF–DCN synapses, and a homosynaptic learning rule at PC–DCN synapses. These were evaluated against a baseline network in which only PF–PC plasticity was active. The three experimental conditions were assessed during the acquisition of a circle-like trajectory. The results show that incorporating plasticity in the DCN improves performance when the initial MF–DCN or PC–DCN synaptic weights are not optimally tuned. Networks with DCN plasticity achieved lower MAE without requiring pre-optimisation of synaptic weights.
Discussion
These results suggest that plasticity in the deep cerebellar nuclei contributes to motor adaptation alongside cortical learning mechanisms. When DCN plasticity is absent, achieving low error requires careful pre-optimisation of synaptic weights to match the robot’s joint dynamics and actuator properties. In contrast, enabling DCN learning allows the cerebellar controller to adapt synaptic weights online, reducing the need for manual parameter tuning. This supports the hypothesis that cerebellar nuclear plasticity plays a functional role in adaptive motor control.
References
[1] Husbands, P., Shim, Y., Garvie, M., Dewar, A., Domcsek, N., Graham, P., ... & Philippides, A. (2021). Recent advances in evolutionary and bio-inspired adaptive robotics: Exploiting embodied dynamics. Applied Intelligence, 51(9), 6467-6496.
[2] Glickstein, M., & Doron, K. (2008). Cerebellum: connections and functions. The Cerebellum, 7(4), 589-594.
[3] Marr, D. (1969). A theory of cerebellar cortex. The Journal of physiology, 202(2), 437-470.
[4] Gao, Z., Van Beugen, B. J., & De Zeeuw, C. I. (2012). Distributed synergistic plasticity and cerebellar learning. Nature Reviews Neuroscience, 13(9), 619-635.
[5] Abadía, I., Naveros, F., Garrido, J. A., Ros, E., & Luque, N. R. (2019). On robot compliance: A cerebellar control approach. IEEE Transactions on Cybernetics, 51(5), 2476-2489.
Acknowledgement
NY was funded by a grant from the Academy of Medical Sciences.
