Locomotion Policy Guided Traversability Learning using Volumetric Representations of Complex Environments

Abstract: Despite the progress in legged robotic locomotion, autonomous navigation in unknown environments remains an open problem. Ideally, the navigation system utilizes the full potential of the robots' locomotion capabilities while operating within safety limits under uncertainty. The robot must sense and analyze the traversability of the surrounding terrain, which depends on the hardware, locomotion control, and terrain properties. It may contain information about the risk, energy, or time consumption needed to traverse the terrain. To avoid hand-crafted traversability cost functions we propose to collect traversability information about the robot and locomotion policy by simulating the traversal over randomly generated terrains using a physics simulator. Thousand of robots are simulated in parallel controlled by the same locomotion policy used in reality to acquire 57 years of real-world locomotion experience equivalent. For deployment on the real robot, a sparse convolutional network is trained to predict the simulated traversability cost, which is tailored to the deployed locomotion policy, from an entirely geometric representation of the environment in the form of a 3D voxel-occupancy map. This representation avoids the need for commonly used elevation maps, which are error-prone in the presence of overhanging obstacles and multi-floor or low-ceiling scenarios. The effectiveness of the proposed traversability prediction network is demonstrated for path planning for the legged robot ANYmal in various indoor and natural environments.

• The paper introduces a novel method for robot traversability estimation using 3D volumetric representations of complex environments, moving beyond traditional 2.5D maps.
• Data collection for traversability learning is achieved through extensive physics-based simulation, generating over 57 years equivalent of locomotion experience without manual cost functions.
• A sparse 3D CNN is employed for real-time traversability prediction, validated through deployment on a legged robot in diverse real-world indoor and natural settings.

Locomotion Policy Guided Traversability Learning using Volumetric Representations of Complex Environments

This paper presents a novel approach to autonomous navigation for legged robots, focusing on traversability estimation using volumetric representations. The research addresses a key challenge in robotics: enabling autonomous navigation in unknown and complex environments while fully leveraging the robot's locomotion capabilities within safety constraints.

Key Contributions

1. Volumetric Representation for Traversability Estimation: The study introduces a method that uses a 3D voxel-occupancy map to represent the environment, moving away from traditional 2.5D elevation maps that often fail in complex scenarios with overhanging obstacles and multi-floor settings. This 3D representation captures geometric information, helping avoid obstacles and steep slopes, which is critical for path planning in intricate terrains.
2. Physics-based Simulation for Data Collection: The paper proposes collecting traversability data through simulations over randomly generated terrains using a physics simulator. This method simulates thousands of robots executing the same locomotion policy as used in real-world scenarios, thereby accumulating an equivalent of 57 years of real-world locomotion experience. This innovative approach eliminates the need for hand-crafted traversability cost functions.
3. Sparse 3D CNN for Real-time Traversability Prediction: A sparse convolutional neural network (CNN) is trained to predict traversability costs tailored to the deployed locomotion policy. By focusing on sparsity, the network efficiently processes input data and operates in real-time, making it suitable for deployment in real-world environments such as forests and indoor settings.

Experimental Validation

The paper validates its approach by deploying the system on the legged robot ANYmal in a variety of indoor and natural environments. The experiments demonstrate successful path planning, supporting the practical applicability of the proposed traversability prediction network. The system effectively identifies traversable regions, accounting for various configurations and motion commands, which is particularly advantageous in complex environments with navigation challenges, like under desks and in narrow corridors.

Implications and Future Directions

The implications of this research extend to improving the robustness and efficiency of autonomous navigation systems. The use of a volumetric representation for traversability estimation could inform the development of more advanced navigation algorithms capable of operating in environments previously deemed challenging due to their geometric complexity. Furthermore, the success in using a dense, simulation-based data gathering method opens avenues for training AI systems on diverse and comprehensive datasets without extensive real-world trials.

Looking to the future, this work could inspire improvements in sim-to-real transfer techniques, ensuring that simulations and real-world applications align closely to enhance predictability and reliability in robot navigation. Additionally, integrating semantic information from visual inputs could further refine traversability estimations by assessing terrain properties such as friction.

Overall, this paper solidifies the foundation for more resilient locomotion policies in robotics, contributing significantly to the field's understanding of traversability in 3D environments. Such advancements are vital for the deployment of autonomous robots in exploration and operational tasks across diverse sectors.