"Recent advancements in neuromorphic computing have enabled significant progress in dynamic motion recognition, yet distinguishing high-speed and low-speed movements remains computationally challenging due to the limited dynamic range of conventional CMOS technology. Our work introduces a SnS₂-based in-sensor reservoir computing device that leverages tunable multi-timescale optoelectronic dynamics to address this bottleneck," explained study author Linfeng Sun, a professor at Beijing Institute of Technology. "The device's conductance under light stimulation is governed by the trapping and recombination of photogenerated carriers at inherent defect states, enabling flexible responses to varying illumination durations—from rapid photodetection for high-speed inputs to sustained synaptic-like behavior for slow-motion analysis."
The SnS₂ memristor, fabricated via chemical vapor deposition (CVD), exhibits exceptional optoelectronic properties: a responsivity of 208.9 A/W, external quantum efficiency of 733.98%, and detectivity of 1.1 × 10¹¹ Jones. "By modulating light intensity according to motion speed—low intensity for fast movements and high intensity for slow ones—we achieve hardware-level adaptation of temporal dynamics," said co-author Zhongrui Wang. "This eliminates the saturation and signal loss issues in fixed-response systems."
In motion recognition tasks using the Weizmann dataset, the in-sensor reservoir achieved 100% accuracy in classifying running, side jumping, and walking across varying velocities. The system's computational efficiency surpasses conventional networks like LSTM with only 20,739 parameters. "Our approach collocates sensing, memory, and processing in a single device, avoiding the von Neumann bottleneck," Sun emphasized. "While scalability to high-resolution datasets like UCF101 requires further work, this SnS₂-based platform offers a promising pathway for energy-efficient, real-time edge intelligence in autonomous systems and robotics."
Authors of the paper include Ruitong Bie, Xi Chen, Zhe Yang, Dong An, Yifei Yu, Qianyu Zhang, Ce Li, Zirui Zhang, Dingchen Wang, Jichang Yang, Songqi Wang, Binbin Cui, Dongliang Yang, Lin Hu, Zhongrui Wang, and Linfeng Sun.
This work was supported by the Beijing Natural Science Foundation (grant no. Z210006) and National Key R&D Plan (2022YFA1405600). Z.W. appreciates the financial support from Hong Kong RGC (nos. 27206321, 17205921, and 17212923), National Natural Science Foundation of China (no. 62122004), and Shenzhen Science and Technology Commission (SGDX2022053011405040). L.H. thanks the support from the National Natural Science Foundation of China (grant no. 12274024).
The paper, "Tunable Neuromorphic Computing for Dynamic Multi-Timescale Sensing in Motion Recognition" was published in the journal Cyborg and Bionic Systems on Sep 30, 2025, at DOI: 10.34133/cbsystems.0412.
