Sim-to-Real Fall Detection for ICU Beds

Sat, 09 May 2026 00:00:00 +0000

Overview

Falls from ICU beds are a serious and persistent patient-safety concern, but real fall events are rare, dangerous to reproduce, and ethically difficult to collect at scale. This project asks whether that gap can be closed with simulation: we build a physics-based bed environment that synthesizes a wide range of in-bed postures and movements, including both falls and ordinary non-fall motion, and uses the resulting episodes to design a continuous fall-risk label rather than a simple fall/no-fall flag.

Training directly on raw simulated video or dense sensor output is often too heavy for the embedded boards found in real ICU settings. To keep the model deployable, we represent each posture with a compact set of body joint coordinates and train a lightweight predictor on top of that representation, so the resulting fall-risk model stays small and fast enough to run in real time on constrained hardware.

Because the model is trained in simulation, a central part of this work is understanding how well it transfers to the real world. We study the gap between simulated dynamics and real-world video, examining where risk predictions stay reliable and where simulation-only training falls short, to guide how much real-world data and calibration a deployed system would still need.

This project’s methodology is detailed in our paper, Fall-from-Bed Risk Prediction Using Physics-Based Bed Simulation, published in MDPI Sensors (2026).

Neuromorphic Control of Robotic Manipulators

Mon, 15 Jan 2024 00:00:00 +0000

Abstract

Our research at the Human Brain Neurocomputing Platform Research Center focuses on developing energy-efficient and biologically-inspired robotic control systems using spiking neural networks (SNNs).

Unlike traditional artificial neural networks (ANNs) that suffer from high energy consumption and real-time processing limitations, SNNs mimic biological neurons’ spike-based information processing mechanisms, offering superior energy efficiency and temporal dynamics suitable for real-time applications.

We are currently developing a neuromorphic hardware-friendly reward-modulated spike timing-dependent plasticity (R-STDP) framework integrated with twin delayed deterministic policy gradient (TD3) reinforcement learning algorithms for 3-degree-of-freedom robotic arm control.

This approach simplifies complex neuromorphic learning schemes while enabling on-chip online learning capabilities with significantly reduced computational overhead compared to traditional backpropagation methods. Our work aims to bridge the gap between biological neural computation and practical robotic applications, demonstrating that SNN-based systems can achieve robust adaptive control while maintaining the ultra-low power consumption characteristics essential for next-generation autonomous systems and edge computing applications.

Key Features

Energy Efficiency: Ultra-low power consumption through spike-based processing
Biologically Inspired: Mimics natural neural computation
Real-time Control: Suitable for time-critical robotic applications
On-chip Learning: R-STDP enables online learning on neuromorphic hardware

Keywords

Spiking Neural Networks (SNNs)
Reward-modulated STDP (R-STDP)
Twin Delayed Deterministic Policy Gradient (TD3)
3-DOF Robotic Arm Control

Description

Our research at the Human Brain Neurocomputing Platform Research Center focuses on developing energy-efficient and biologically-inspired robotic control systems using spiking neural networks (SNNs). Unlike traditional artificial neural networks (ANNs) that suffer from high energy consumption and real-time processing limitations, SNNs mimic biological neurons’ spike-based information processing mechanisms, offering superior energy efficiency and temporal dynamics suitable for real-time applications. We are currently developing a neuromorphic hardware-friendly reward-modulated spike timing-dependent plasticity (R-STDP) framework integrated with twin delayed deterministic policy gradient (TD3) reinforcement learning algorithms for 3-degree-of-freedom robotic arm control. This approach simplifies complex neuromorphic learning schemes while enabling on-chip online learning capabilities with significantly reduced computational overhead compared to traditional backpropagation methods. Our work aims to bridge the gap between biological neural computation and practical robotic applications, demonstrating that SNN-based systems can achieve robust adaptive control while maintaining the ultra-low power consumption characteristics essential for next-generation autonomous systems and edge computing applications.

Physical AI & Sim-to-Real | BCML Lab.