Abstract
We present a novel outdoor navigation algorithm to generate stable and efficient actions to navigate a robot to reach a goal. We use a multi-stage training pipeline and show that our approach produces policies that result in stable and reliable robot navigation on complex terrains. Based on the Proximal Policy Optimization (PPO) algorithm, we developed a novel method to achieve multiple capabilities for outdoor navigation tasks, namely alleviating the robot's drifting, keeping the robot stable on bumpy terrains, avoiding climbing on hills with steep elevation changes, and avoiding collisions. Our training process mitigates the reality (sim-to-real) gap by introducing generalized environmental and robotic parameters and training with rich features of Lidar perception in a high-fidelity Unity simulator. We evaluate our method in both simulation and real world environments using Clearpath Husky and Jackal robots. Further, we compare our method against the state-of-the-art approaches and observe that, in the real world it improves stability by at least 30.7% on uneven terrains, reduces drifting by 8.08% and decreases the elevation changes by 14.75%.
Introduction
Navigating mobile robots in outdoor environments presents significant challenges compared to controlled indoor settings. Unstructured terrains—characterized by slippery mud, wet leaves, bumpy grass, and steep hills—introduce unpredictable physics that often lead to navigation failures. Specifically, current systems struggle with the "sim-to-real gap," where simulators fail to capture complex friction and vibration responses. Furthermore, robots often face a trade-off between speed and stability; excessive speed on rough ground can lead to mechanical damage or tipping. Finally, "drifting" on low-friction surfaces causes robots to deviate from intended paths, increasing collision risks. AdaptiveON addresses these issues by focusing not just on traversability, but on the reliability and safety of the navigation process.
Methodology
The AdaptiveON framework utilizes a multi-stage training pipeline based on the Proximal Policy Optimization (PPO) algorithm. The architecture is modular, integrating perception, adaptation, and stability control.
The system consists of four primary components:
- Perception Models: These process 16-channel Lidar data and point-cloud-based elevation maps to identify obstacles and terrain geometry.
- Adaptive Model: This module analyzes a short history of the robot’s previous states and actions. By comparing commanded actions with actual motion, it implicitly estimates environmental parameters like friction and terrain restitution.
- Stable Model: This component monitors vibrations (roll and pitch angular velocities). It generates a velocity threshold to throttle the robot on bumpy terrain, ensuring mechanical stability.
- Policy Model: The final decision-maker that integrates inputs from all modules to produce linear and angular velocity commands.
A key innovation is the use of "history as a sensor," where the Adaptive Model uses past motion to recognize slippery surfaces and adjust future actions. Additionally, the Stable Model acts as a mechanical safeguard, preventing the policy from becoming overconfident on dangerous ground.
Experimental Setup
The system was trained using a curriculum-style approach across two simulators: RaiSim, for customizing physics like friction and bumps, and Unity, for high-fidelity Lidar simulation.
The evaluation involved five distinct simulation stages, including adaptive scenarios with varying friction, uneven terrain with Perlin noise-generated bumps, static obstacle avoidance, and steep hills. Real-world deployments were conducted on Clearpath Jackal and Husky robots at the University of Maryland, covering bumpy grass, slippery rocks, and steep hills.
Results and Discussion
In real-world tests, AdaptiveON demonstrated significant improvements over state-of-the-art methods such as Terp, CrowdSteer, and Ego-graph:
- Stability: Vibration was reduced by at least 30.7% on uneven terrains.
- Drift Reduction: Navigation drift on slippery surfaces decreased by 8.08%.
- Elevation Management: The robot actively chose paths with 14.75% gentler slopes, effectively avoiding hazardous hills.
- Generalization: The system successfully transferred from simulation to two different robot platforms without extensive real-world fine-tuning.
The multi-stage curriculum was found to be essential, as training functionalities consecutively prevented conflicting objectives from interfering during the learning process.
Conclusion
AdaptiveON provides a robust framework for stable outdoor navigation. While the system significantly improves reliability, current limitations include a slight reactive latency in the Stable Model and challenges in perceiving features like tall grass. Future work will focus on reducing this latency and improving the decoupling of drift-prevention and stability functions for even more complex scenarios.
Summary updated