Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Navigating the Unknown: Uncertainty-Aware Compute-in-Memory Autonomy of Edge Robotics

Nastaran Darabi, Priyesh Shukla, Dinithi Jayasuriya, Divake Kumar, Alex C. Stutts, Amit Ranjan Trivedi

TL;DR

The paper tackles energy-constrained, uncertainty-aware pose estimation for insect-scale edge drones operating in dynamic environments. It advances a co-design approach that integrates Compute-in-Memory (CIM) hardware with Bayesian and probabilistic inference, applying CIM to both particle filtering for localization and MC-Dropout-based variational inference for visual odometry. Key contributions include CIM-driven HMGM-based likelihood computation on inverter arrays for ultra-low-power localization, CIM-accelerated variational inference with compute reuse for Bayesian neural networks, and SRAM-embedded RNGs enabling efficient dropout sampling, achieving high energy efficiency while expressing predictive uncertainty. The work demonstrates that robust, uncertainty-aware edge navigation can be realized within stringent area/power budgets, and suggests future MC-free uncertainty frameworks to further reduce hardware demands.

Abstract

This paper addresses the challenging problem of energy-efficient and uncertainty-aware pose estimation in insect-scale drones, which is crucial for tasks such as surveillance in constricted spaces and for enabling non-intrusive spatial intelligence in smart homes. Since tiny drones operate in highly dynamic environments, where factors like lighting and human movement impact their predictive accuracy, it is crucial to deploy uncertainty-aware prediction algorithms that can account for environmental variations and express not only the prediction but also confidence in the prediction. We address both of these challenges with Compute-in-Memory (CIM) which has become a pivotal technology for deep learning acceleration at the edge. While traditional CIM techniques are promising for energy-efficient deep learning, to bring in the robustness of uncertainty-aware predictions at the edge, we introduce a suite of novel techniques: First, we discuss CIM-based acceleration of Bayesian filtering methods uniquely by leveraging the Gaussian-like switching current of CMOS inverters along with co-design of kernel functions to operate with extreme parallelism and with extreme energy efficiency. Secondly, we discuss the CIM-based acceleration of variational inference of deep learning models through probabilistic processing while unfolding iterative computations of the method with a compute reuse strategy to significantly minimize the workload. Overall, our co-design methodologies demonstrate the potential of CIM to improve the processing efficiency of uncertainty-aware algorithms by orders of magnitude, thereby enabling edge robotics to access the robustness of sophisticated prediction frameworks within their extremely stringent area/power resources.

Navigating the Unknown: Uncertainty-Aware Compute-in-Memory Autonomy of Edge Robotics

TL;DR

The paper tackles energy-constrained, uncertainty-aware pose estimation for insect-scale edge drones operating in dynamic environments. It advances a co-design approach that integrates Compute-in-Memory (CIM) hardware with Bayesian and probabilistic inference, applying CIM to both particle filtering for localization and MC-Dropout-based variational inference for visual odometry. Key contributions include CIM-driven HMGM-based likelihood computation on inverter arrays for ultra-low-power localization, CIM-accelerated variational inference with compute reuse for Bayesian neural networks, and SRAM-embedded RNGs enabling efficient dropout sampling, achieving high energy efficiency while expressing predictive uncertainty. The work demonstrates that robust, uncertainty-aware edge navigation can be realized within stringent area/power budgets, and suggests future MC-free uncertainty frameworks to further reduce hardware demands.

Abstract

This paper addresses the challenging problem of energy-efficient and uncertainty-aware pose estimation in insect-scale drones, which is crucial for tasks such as surveillance in constricted spaces and for enabling non-intrusive spatial intelligence in smart homes. Since tiny drones operate in highly dynamic environments, where factors like lighting and human movement impact their predictive accuracy, it is crucial to deploy uncertainty-aware prediction algorithms that can account for environmental variations and express not only the prediction but also confidence in the prediction. We address both of these challenges with Compute-in-Memory (CIM) which has become a pivotal technology for deep learning acceleration at the edge. While traditional CIM techniques are promising for energy-efficient deep learning, to bring in the robustness of uncertainty-aware predictions at the edge, we introduce a suite of novel techniques: First, we discuss CIM-based acceleration of Bayesian filtering methods uniquely by leveraging the Gaussian-like switching current of CMOS inverters along with co-design of kernel functions to operate with extreme parallelism and with extreme energy efficiency. Secondly, we discuss the CIM-based acceleration of variational inference of deep learning models through probabilistic processing while unfolding iterative computations of the method with a compute reuse strategy to significantly minimize the workload. Overall, our co-design methodologies demonstrate the potential of CIM to improve the processing efficiency of uncertainty-aware algorithms by orders of magnitude, thereby enabling edge robotics to access the robustness of sophisticated prediction frameworks within their extremely stringent area/power resources.
Paper Structure (11 sections, 2 equations, 3 figures)

This paper contains 11 sections, 2 equations, 3 figures.

Table of Contents

  1. Introduction
  2. Compute-in-Memory Particle Filtering-based Drone Localization
  3. Overview of Particle Filtering-based Localization
  4. Posterior Density Estimation with an Inverter Array
  5. Compute-in-Memory Particle Filtering
  6. Compute-in-Memory Uncertainty-Expressive Bayesian Visual Odometry
  7. Overview of Visual Odometry (VO)
  8. Bayesian and Variational Inference of Neural Networks
  9. Compute-in-Memory Monte-Carlo Dropout-based
  10. Uncertainty Expressive Visual Odometry (VO)
  11. Conclusions and Future Works

Figures (3)

  • Figure 1: Various sources of uncertainties arise in typical autonomous systems such as the inability to perceive the application domain reliably, discrepancies in the trained and applied environment, and sources of non-idealities such as noise and process variability impacting the reliability of processing.
  • Figure 2: Compute-in-memory-based ultra-low-power Monte-Carlo localization framework for insect-scale drones.(a) CIM architecture employs columns of 6-T inverters to leverage their switching current for likelihood evaluations. (b) Gaussian-like switching current of 6-T inverters. (c) Contour and (d) surface plot of switching current at varying input voltages showing rectilinear tails of the implemented function whereas a Gaussian function has elliptical tails. On RGB-D dataset (e), (f)-(h) show the comparison of localization steps between the proposed co-designed HMGM-based approach vs. conventional GMM-based approach. (i) Comparison of energy between an 8-bit digital processor and 4-bit HMGM inverter array-based CIM. [Reproduced from our prior work shukla2021ultralow under IEEE's open-source license agreement.]
  • Figure 3: Compute-in-memory-based acceleration of variational inference of neural networks:(a) An SRAM-based CIM macro combines storage and Bayesian inference (BI) functionalities. The inset figure illustrates an 8T SRAM cell featuring both storage and product ports. The CIM incorporates a randomly generated dropout bit generator, enabling MC-Dropout inference. (b) CIM-embedded dropout bit generator. MC-Dropout for indoor (RGB-D dataset) drone localization and trajectory estimation in (c) X-Y, (d) Y-Z, and (e) X-Z position coordinates. We also compare these results with deterministic network configurations under various inference conditions. (f) The correlation between uncertainty (variance) and the error in the pose estimates. [Reproduced from shukla2022mc with IEEE's permission]