Why Kompass?#
Robot tasks involve the interaction of multiple functional units, making it crucial to focus on the overall system design rather than just the design of individual components. This is precisely the case for Autonomous Navigation, where multiple parts come together to form a cohesive system.
An autonomously navigation agent should be able to adapt to changing environmental conditions and deploy different behaviours on the fly to meet the changes incurring in the scene. This seemingly simple requirement becomes challenging to implement even in static environments. This was a direct conclusion of the BARN challenge taking place at ICRA every year since 2021:
“… while it is worthwhile to extend navigation research in directions orthogonal to metric navigation, the community should also not overlook the problems that still remain in this space, especially when robots are expected to be extensively and reliably deployed in the real world.”
—Lessons learned from The BARN Challenge at ICRA 2022, full article
“All teams adopted a hybrid paradigm in terms of a finite-state-machine setup, which requires different components to address different situations in the obstacle courses, … Such a pragmatic practice suggests that a single stand-alone approach that is able to address all variety of obstacle configurations all together is still out of our reach.”
—Lessons learned from The 3rd BARN Challenge at ICRA 2024, full article
Currently, the only other full-system navigation solution in the open-source community is Nav2, so why did we decide to create Kompass?
Adaptive event-driven design to the core#
Kompass is designed as per the specification of open event-driven software standard to dynamically respond to real-time changes in the environment, the robots internal state, or the task assigned. Mobile robots working in interaction with messy world dynamics require to adapt on the fly and ensure stable performance in unpredictably changing conditions. Handling external events is built in the core of Kompass so the user can easily design a complete system capable of reconfiguring itself on the fly when an event is perceived in the world or a task is issued to the robot, making it more adaptive and robust.
For example, with Kompass the user can easily design a system that utilizes one planning system when the robot is out on the road or another controller when its inside the building or yet another when its close to its docking station, while easily configuring fallback conditions for each for those components. This approach of providing event driven control over the stack itself, makes it much more flexible to implement a comprehensive autonomous navigating agent which can operate in multiple scenarios. This is in contrast to other approaches like defining behaviour trees where the changes in the stack are based on internal state of the robot alone (and that too with a rather complicated API).
Machine learning models as first class citizens#
The event driven stack control allows Kompass to utilize machine learning models unlike any other navigation stack. External events in Kompass can be driven by outputs of machine learning models interpreting sensor data or user commands, which means the entire stack becomes reconfigurable based on ML model outputs. This goes beyond well established scenarios of visual navigation.
As an example consider the scenario where the robot observes a certain number of humans in its environment and switches from a path following controller like Pure Pursuit to a predictive controller like Timed Elastic Bands or even better a human-aware controller like Human Aware TEB. The same goes for utilizing outputs of VLMs, which can answer more abstract aggregate perception questions like ‘Whether the robot is inside or outside?’, to change control behaviour on the fly.
To get more ideas about utilizing machine learning models with Kompass and create intelligent embodied agents, check out integration of Kompass with ROS Agents.
Ease of use and intuitive API while remaining within the ROS ecosystem#
Kompass provides an approachable and intuitive interface for creating navigation systems. This is made possible with a pythonic API using ROS Sugar. A fairly sophisticated navigation system can be configured in one simple python script where the user can configure the stack component by component and configure the system level behavior by defining all the required events, their consequent actions and per component fallback behaviors. The user also has a choice to provide component level parameters in YAML files. Furthermore, the project is structured so that core algorithms are implemented in the Kompass Core package, which provides a pure python interface while the underlying algorithms can be implemented in both python for quick prototyping and in C/C++ for performance optimization in production environments.
Modular architecture and easy extensibility#
The architecture of Kompass to separate the core algorithms in Kompass Core package, away from its ROS primitives also serves to simplify upgrades and minimizes the risk of breaking changes between different ROS versions and distributions. It also allows for seamless integration of additional planning and control algorithms, as well as managing integration with specialized third party libraries. Most importantly, Kompass has been designed to be extendable by the community as a unified place to contribute additional plan and control algorithms that can be useful for mobile robots.