Jean-François Tremblay,Julie Alhosh,Louis Petit,Faraz Lotfi,Lara Landauro,David Meger
Abstract:Autonomous robots navigating in off-road terrain like forests open new opportunities for automation. While off-road navigation has been studied, existing work often relies on clearly delineated pathways. We present a method allowing for long-range planning, exploration and low-level control in unknown off-trail forest terrain, using vision and GPS only. We represent outdoor terrain with a topological map, which is a set of panoramic snapshots connected with edges containing traversability information. A novel traversability analysis method is demonstrated, predicting the existence of a safe path towards a target in an image. Navigating between nodes is done using goal-conditioned behavior cloning, leveraging the power of a pretrained vision transformer. An exploration planner is presented, efficiently covering an unknown off-road area with unknown traversability using a frontiers-based approach. The approach is successfully deployed to autonomously explore two 400 meters squared forest sites unseen during training, in difficult conditions for navigation.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is to achieve long - distance navigation of autonomous robots in off - road terrains. Specifically, the author proposes a method that enables the robot to perform long - distance planning, exploration, and low - level control in unknown forest terrains, using only vision and GPS. This method focuses particularly on navigation without a clear path, that is, in complex environments such as forests, the robot needs to be able to identify and bypass obstacles, such as fallen trees, rivers, etc., and reach the target location safely.
### Main contributions of the paper:
1. **Visual traversability learning method**: By leveraging the knowledge of human experts, predict whether there is a safe path to the target in the image.
2. **Topological mapping and exploration framework**: Construct a topological graph, where edges represent the traversability determined by the above - mentioned model.
3. **Behavior cloning controller**: Capable of navigating in very rugged terrains, including landslides, obstacles, and narrow spaces.
4. **Extensive field evaluation**: Evaluated in two unseen forest terrains, with a total area of 800 square meters and a travel distance of 700 meters under extremely harsh conditions.
### Method overview:
- **Topological mapping and planning**: Represent the environment using a topological graph, where nodes are GPS waypoints and edges represent the reachability between adjacent nodes. The robot explores the environment by selecting the nearest reachable frontier node.
- **Traversability analysis**: Define image - level traversability and train a network through supervised learning to predict whether there is a safe path in the image. Edge - level traversability is determined through logical judgment.
- **Goal - conditioned behavior cloning controller**: Learn the control strategy through behavior cloning, enabling the robot to navigate from one node to another. The controller outputs linear velocity and steering rate based on the current context (image, speed, past commands) and the visible nearby target.
- **Exploration planning**: Use a frontier - based exploration algorithm, select the nearest time - reachable frontier node for exploration, and plan the path through the A* algorithm.
### Experimental setup:
- **Robot platform**: Use the Clearpath Husky robot, equipped with three OAK - D wide - angle lens cameras, an IMU, and an RTK GPS system.
- **Experimental area**: Conduct experiments in Maisonneuve Park in Montreal and the Gault Nature Reserve at McGill University.
- **Training and deployment methods**: Include manual data collection, manual traversability labeling, neural network training, semi - autonomous deployment and data collection, and test deployment.
### Results:
- **Traversability network**: On test sites A and B, the classification results of the traversability network were compared with the results of manual labeling, showing high accuracy.
- **Behavior cloning controller**: During the test, the controller performed well, but there were still a small number of intervention cases, mainly because the robot hit small trees instead of bypassing them.
- **Planning and overall system**: Successfully explored the two test areas, visited each reachable cell of 2 meters by 2 meters, and the total travel distance was 700 meters. There were a total of 25 interventions, mainly due to errors in the controller or traversability analysis.
### Conclusion:
This paper proposes an algorithm framework for achieving autonomous navigation in off - road environments. Through sparse, traversability - aware topological mapping and a simple behavior cloning controller, two unseen forest areas were successfully explored. Future work will focus on vision - based frontier selection and actively collecting traversability data to reduce the need for labeling.