ConfigBot: Adaptive Resource Allocation for Robot Applications in Dynamic Environments

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

Introduction

Modern robotics platforms are essentially mobile, resource-constrained edge computing systems. Unlike cloud-based environments that benefit from horizontal scaling, a robot operates within the rigid boundaries of its onboard CPU, GPU, and memory. Traditionally, developers have managed these constraints through static configurations or significant over-provisioning to ensure system stability. However, robot workloads are inherently dynamic; a robot navigating a sparse environment requires far fewer resources than one operating in a crowded, obstacle-rich space.

The authors identify a critical "throttling trap" in standard Linux schedulers, such as the Completely Fair Scheduler (CFS). While these schedulers prioritize fairness, they often fail in robotics contexts where delays in "core" services (e.g., localization or obstacle avoidance) can lead to catastrophic physical failures. Simply limiting a process's CPU usage often results in thread accumulation, which paradoxically increases latency and destabilizes the system.

The ConfigBot Framework

ConfigBot addresses these challenges by treating resource allocation as a constrained multi-objective optimization problem. The framework is built upon three primary pillars:

  1. Service Categorization: Applications are divided into "Core" and "Non-Core" services. Core services (e.g., navigation) are treated as hard constraints that must meet specific performance thresholds. Non-core services (e.g., telemetry, secondary perception) are treated as objectives to be maximized within the remaining resource budget.
  2. The Adaptor Abstraction: A key innovation of ConfigBot is the introduction of "adaptors" at the Robot Operating System (ROS) layer. Rather than relying solely on OS-level cgroups to throttle CPU cycles, these adaptors manage data flow by dropping or rate-limiting messages between nodes. This approach reduces the actual computational workload, preventing the "thread pile-up" associated with raw CPU starvation.
  3. Bayesian Optimization (BO): To navigate the complex, high-dimensional space of configuration parameters (including CPU limits and adaptor rates), ConfigBot employs Bayesian Optimization. By profiling the system in short 5-second bursts, the framework learns the performance impact of various settings and iteratively identifies the Pareto optimal configuration for the current environmental context.

Continuous Monitoring and Implementation

ConfigBot utilizes eBPF and rosmaster monitors to provide real-time visibility into system performance. This monitoring allows the framework to detect environmental shifts or workload changes—such as the activation of a new application—and trigger a re-optimization cycle if the current configuration no longer satisfies the developer's requirements. Crucially, by integrating adaptors into the ros_comm library, ConfigBot remains application-transparent, requiring no modifications to the original application code.

Experimental Evaluation

The framework was evaluated across three distinct robotic platforms: a Boston Dynamics Spot (equipped with an NVIDIA AGX Orin), a Clearpath Jackal (Intel i7 8-core), and a custom mobile Cobot (Intel NUC). The testing involved various navigation stacks (Basic-NAV, Intermed-NAV, Terrain-NAV) and non-core applications like YOLO-based object detection and web telemetry.

The results demonstrate significant improvements over default OS configurations:

  • Resource Efficiency: On the Spot robot, the "Basic-NAV" stack met all performance requirements using 4x fewer CPU resources than the default configuration.
  • System Stability: ConfigBot achieved a near-100% constraint satisfaction rate. In contrast, the default Linux configuration frequently led to "stumbles" or pauses due to resource contention.
  • Performance Gains: Non-core application performance improved by 49% to 193% while maintaining the integrity of core navigation services.

Discussion and Future Work

A central insight of this research is the "Throttling Paradox": restricting CPU time in multithreaded robotics applications can be counterproductive. Instead, controlling the data flow (input) is a more effective lever for maintaining system stability. Furthermore, the experiments confirmed that there is no "one size fits all" configuration; optimal settings are highly context-dependent, validating the need for dynamic re-learning.

Despite its success, ConfigBot has limitations. It currently requires manual specification of performance metrics and targets (e.g., requiring navigation to run at >35Hz). Future iterations could explore inferring these specifications from high-level task descriptions. Additionally, while the current reactive model is effective, moving toward a proactive model using Reinforcement Learning (RL) could allow the system to anticipate bottlenecks and adjust configurations before performance drops.

Summary updated

BibTeX
@inproceedings{dwivedula2025configbot,
  author = {Rohit Dwivedula and Sadanand Modak and Aditya Akella and Joydeep Biswas and Daehyeok Kim and Christopher J. Rossbach},
  title = {{ConfigBot: Adaptive Resource Allocation for Robot Applications in Dynamic Environments}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2025},
  doi = {10.1109/iros60139.2025.11245932},
}
Figure