Unified Autonomy Stack lets drones and legged robots navigate smoke-filled, GPS-denied areas
Open-source stack fuses LiDAR, radar, and vision for resilient robot autonomy in extreme conditions
A team of researchers led by Mihir Dharmadhikari and Kostas Alexis introduced and open-sourced the Unified Autonomy Stack, a comprehensive system-level solution for generalizable robot autonomy. The architecture is built around three synergistic modules: multi-modal perception, multi-behavior planning, and multi-layered safe navigation. These modules work together to deliver mission autonomy across diverse robot morphologies, including both aerial (rotorcraft) and ground (legged) robots.
The perception module fuses data from LiDAR, radar, vision, and inertial sensors to enable robust localization and mapping through factor graph-based fusion, as well as semantic scene understanding. The planning module uses sampling-based techniques adaptive across spatial scales for motion and informative path planning. The navigation module employs a multi-layered approach: planning on the online reconstructed map, deep learning-driven exteroceptive policies, and last-resort safety filters using control barrier functions. This layered design ensures safe behavior even in unknown, perceptually-degraded, or cluttered environments.
The stack has been extensively field-tested in challenging conditions, including GNSS-denied areas, self-similar environments (e.g., tunnels), and smoke-filled settings with complex geometries and high obstacle clutter. It enables behaviors such as safe navigation into unknown regions, exploration of complex environments, object discovery, and efficient inspection planning. The researchers published a 35-page paper with 22 figures and 8 tables detailing the methodology and experimental results. They also made the full implementation, documentation, validation datasets, and a video overview available to facilitate adoption by the robotics community.
- Integrates multi-modal perception (LiDAR, radar, vision, inertial) with factor graph-based localization and semantic scene understanding.
- Uses a triple-layer navigation scheme: online map-based planning, deep learning policies, and control barrier functions as safety filters.
- Validated on both aerial (rotorcraft) and legged robots in GNSS-denied, smoke-filled, and clutter-heavy environments.
Why It Matters
This open-source stack provides a robust blueprint for deploying robots in GPS-denied, low-visibility environments critical for search/rescue and industrial inspection.