ConfigBot: Adaptive Resource Allocation for Robot Applications in Dynamic Environments

Rohit Dwivedula, Sadanand Modak, Aditya Akella, Joydeep Biswas, Daehyeok Kim, Christopher J. Rossbach · 2025

Read ConfigBot: Adaptive Resource Allocation for Robot Applications in Dynamic Environments by Rohit Dwivedula, Sadanand Modak, Aditya Akella, Joydeep Biswas, Daehyeok Kim, Christopher J. Rossbach (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2501-10513.txt.

Introduction

Rohit Dwivedula and co-authors present ConfigBot, a system that retunes robots as conditions change. Their 2025 paper notes that service robots juggle tasks like navigation and perception on limited onboard compute. Fixed system settings waste resources or cause unstable behavior when loads shift. Their system profiles running apps and tweaks settings on its own to meet a set performance target.

Tests ran on several real robots, each with its own software stack and needs. Some needs changed with context, like place or task phase. ConfigBot held stability while using resources well across these trials. The authors argue that automatic tuning suits robot fleets in shifting settings. It replaces hand-set configs and over-built hardware margins.

Catalog record

CREATORRohit Dwivedula, Sadanand Modak, Aditya Akella, Joydeep Biswas, Daehyeok Kim, Christopher J. Rossbach
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE38 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2501-10513

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

ConfigBot: Adaptive Resource Allocation for Robot Applications in Dynamic Environments Rohit Dwivedula, Sadanand Modak, Aditya Akella, Joydeep Biswas, Daehyeok Kim, Christopher J. Rossbach Abstract The growing use of service robots in dynamic environments requires flexible management of on-board compute resources to optimize the performance of diverse tasks such as navigation, localization, and perception. Current robot deployments often rely on static OS configurations and system over-provisioning. However, they are suboptimal because they do not account for variations in resource usage. This results in poor system-wide behavior such as robot instability or inefficient resource use. This paper presents ConfigBot , a novel system designed to adaptively reconfigure robot applications to meet a predefined performance specification by leveraging runtime profiling and automated configuration tuning . Through experiments on multiple real robots, each running a different stack with diverse performance requirements, which could be context -dependent, we illustrate ConfigBot ’s efficacy in maintaining system stability and optimizing resource allocation. Our findings highlight the promise of automatic system configuration tuning for robot deployments, including adaptation to dynamic changes. I Introduction TABLE I : Common robot apps: typical resource needs, execution frequency and performance metrics Application Frequency Resources Used Performance Metrics # of topics published to subscribed to LiDAR processing 40 - 60 Hz I/O, CPU Latency, Density 12 2 Navigation 20 - 40 Hz CPU, GPU No collisions, smooth movements 25 18 Localization 30 - 40 Hz CPU Accuracy, Recovery 4 5 Web dashboard 10 Hz Network Latency/delay, throughput 3 12 Object detection 1-5 Hz CPU, GPU Frame rate, detection accuracy 2 2 Service robots are increasingly being used to assist humans in everyday tasks, with commercial examples including Amazon Astro, EEVE Willow, and Misty the Robot. Currently, these robots rely on vendor-programmed functions, but we believe that future extensibility through third-party applications (apps) will unlock diverse capabilities - much like app stores ( e.g . Google Play Store) did for smartphones. Furthermore, the set of actively running apps can change dynamically—whether through user-driven actions like starting or stopping functionality, installation/removal of apps, or robot personalization per user [ 1 , 2 ] , thereby significantly expanding the range of workloads that can be supported on a given robot. Real-time responsiveness is crucial for robot apps, as delays or outdated data can hamper effective functionality in dynamic environments (§ IV-A ). Ensuring that numerous concurrently running and dynamically changing sets of apps can effectively share the computational resources on robots while meeting their real-time goals is thus an important resource allocation problem. Most consumer-grade robots are inherently resource-constrained, meaning they cannot simply scale resources on demand—unlike cloud servers—to meet apps’ performance needs. This makes the resource allocation problem in robots particularly challenging. Robot apps’ distinct attributes further add to the challenge (§ III-B ). First, robot apps have highly diverse resource requirements, performance targets, and service level objectives (SLOs). Second, robot apps operate in various environments—such as homes, workplaces, and outdoor settings—where resource demands vary significantly due to environmental variations. Third, robots frequently adapt their algorithms based on the environment [ 3 ] , adding to resource demand volatility. Modern operating systems (OSes)–particularly the Linux variants widely used in robotics–fail to account for these context-specific app resource needs due to reliance on simple resource allocation heuristics. A potential workaround is to over-provision resources and manually set static allocations for each app such that it performs adequately across all contexts; however, this is impractical on resource-constrained systems. To address these challenges, we introduce ConfigBot , a novel resource allocation framework for service robots that automatically tailors app resource allocations to dynamic contexts by building on three unique opportunities that derive from properties of robot apps, as described next. First, robot apps involve persistent, long-running tasks like object tracking and sensor processing [ 4 ] . Thus, ConfigBot optimizes resource allocation infrequently targeting stable environments with consistent workloads, and relies on lightweight runtime monitoring to detect significant environmental or workload changes. Further, ConfigBot maintains a library of proven configurations linked to specific environments and quickly reapplies them in familiar contexts, ensuring efficiency with minimal overhead. Second, robot functionality can be divided into core services ( e.g . localization, navigation) and non-core apps . Building on this, ConfigBot formulates resource allocation as an optimization problem, treating core service performance as constraints and non-core app performance as objectives . This reduces decision variables by limiting essential service tuning, and simplifying optimization. Third, many robot apps operate in a data-driven manner, performing computations only when new data arrives. Coarse resource limits ( e.g . setting max CPU usage limits) can trigger adverse effects (§ IV-C ) that degrade the performance of the throttled process needlessly. ConfigBot controls data flows – particularly for non-core apps – via the novel adaptor abstraction (§ V ), gracefully reducing computational demand under resource constraints without disrupting critical functions. ConfigBot uses Bayesian Optimization to tune the values of OS-based system-level ( cgroup ) settings, as well as the thresholds for the proposed ROS-layer adaptors to meet a developer-defined specification of desired robot behavior. Once deployed, ConfigBot learns a configuration for the current workload + context that meets the specification and applies this identified configuration to the system – without requiring code changes to the robot apps/services. We also discuss how ConfigBot monitors for context shifts that could invalidate prior configurations and re-optimizes for new workloads and environments automatically. We evaluate ConfigBot on state-of-the-art robot platforms, and illustrate how configurations our system identified allowed a navigation and obstacle-avoidance system on the Boston Dynamics Spot to run on 4\times fewer CPUs while still meeting developer-specified performance requirements (§ VIII-A ). We then show that our solution generalizes to multiple sets of robot apps (§ VIII-B ), varying environments (§ VIII-D ) and other robots (§ VIII-E ). II Related Work Automated config search. Automated config search has been successful at making web servers [ 5 , 6 ] , databases [ 7 ] , and cloud apps [ 8 ] more efficient. They use bayesian optimization [ 5 ] , bandit algorithms [ 8 ] , causal inference [ 6 ] or combinations of these techniques [ 9 ] to tune config knobs. Cherrypick [ 5 ] , for instance, uses bayesian optimization to efficiently identify the optimal cloud configuration ( i.e . , VM type, number of CPUs, etc) for big data analytics jobs. To the best of our knowledge, robot OS tuning via automated config search has not been attempted so far. Context-dependent navigation. Most work in this area primarily focuses on getting a set of "good" parameters under different environments. APPLD [ 3 ] uses a supervised dataset of predefined contexts ( i.e . , different environments) and leverages it to train a supervised classifier to choose the current context which helps it to select the context-conditioned behavior cloning policy. Another work [ 10 ] tries to dynamically find trajectory optimization weights for a Dynamic Window Approach (DWA) planner for straight-line and U-turn contexts. However, none of the aforementioned works can detect workload changes automatically as they do not profile the system, and so they need a way to classify the contexts a priori. III Background In this section, we describe modern robot hardware

Read the complete text