Quadrupedal robots hold promising potential for applications in navigating cluttered environments with resilience akin to their animal counterparts. However, their floating-base configuration makes them susceptible to real-world uncertainties, presenting substantial challenges in locomotion control. Deep reinforcement learning has emerged as a viable alternative for developing robust locomotion controllers. However, approaches relying solely on proprioception often sacrifice collision-free locomotion, as they require front-foot contact to detect stairs and adapt the gait. Meanwhile, incorporating exteroception necessitates a precisely modeled map observed by exteroceptive sensors over time. This work proposes a novel method for fusing proprioception and exteroception through a resilient multi-modal reinforcement learning framework. The proposed method yields a controller demonstrating agile locomotion on a quadrupedal robot across diverse real-world courses, including rough terrains, steep slopes, and high-rise stairs, while maintaining robustness in out-of-distribution situations.

Monolithic 3D integration of neuron and synapse devices is considered a promising solution for energy-efficient and compact neuromorphic hardware. However, achieving optimal performance in both training and inference remains challenging as these processes require different synapse devices with reliable endurance and long retention. Here, we introduce a decoupling strategy to separate training and inference using monolithically integrated neuromorphic hardware with layer-by-layer fabrication. This 3D neuromorphic hardware includes neurons consisting of a single transistor (1T-neuron) in the first layer, long-term operational synapses composed of a single thin-film transistor with a SONOS structure (1TFT-synapses) in the second layer for inference, and durable synapses composed of a memristor (1M-synapses) in the third layer for training. A 1TFT-synapse, utilizing a charge-trap layer, exhibits long retention properties favorable for inference tasks. In contrast, a 1M-synapse, leveraging anion movement at the interface, demonstrates robust endurance for repetitive weight updates during training. With the proposed hybrid synapse architecture, frequent training can be performed using the 1M-synapses with robust endurance, while intermittent inference can be managed using the 1TFT-synapses with long-term retention. This decoupling of synaptic functions is advantageous for achieving a reliable spiking neural network (SNN) in neuromorphic computing.