Learning robust perceptive locomotion for quadrupedal robots in the wild (2201.08117v1)

Abstract: Legged robots that can operate autonomously in remote and hazardous environments will greatly increase opportunities for exploration into under-explored areas. Exteroceptive perception is crucial for fast and energy-efficient locomotion: perceiving the terrain before making contact with it enables planning and adaptation of the gait ahead of time to maintain speed and stability. However, utilizing exteroceptive perception robustly for locomotion has remained a grand challenge in robotics. Snow, vegetation, and water visually appear as obstacles on which the robot cannot step~-- or are missing altogether due to high reflectance. Additionally, depth perception can degrade due to difficult lighting, dust, fog, reflective or transparent surfaces, sensor occlusion, and more. For this reason, the most robust and general solutions to legged locomotion to date rely solely on proprioception. This severely limits locomotion speed, because the robot has to physically feel out the terrain before adapting its gait accordingly. Here we present a robust and general solution to integrating exteroceptive and proprioceptive perception for legged locomotion. We leverage an attention-based recurrent encoder that integrates proprioceptive and exteroceptive input. The encoder is trained end-to-end and learns to seamlessly combine the different perception modalities without resorting to heuristics. The result is a legged locomotion controller with high robustness and speed. The controller was tested in a variety of challenging natural and urban environments over multiple seasons and completed an hour-long hike in the Alps in the time recommended for human hikers.

Learning Robust Perceptive Locomotion for Quadrupedal Robots in the Wild

The paper "Learning Robust Perceptive Locomotion for Quadrupedal Robots in the Wild" addresses the significant challenge of autonomous locomotion in legged robots, integrating exteroceptive and proprioceptive inputs to enhance performance in complex, real-world environments. This work stands out by focusing on overcoming the limitations of existing proprioceptive-only frameworks, which restrict speed and adaptability due to reliance on tactile feedback alone.

Methodology and Implementation

The researchers present an innovative locomotion controller utilizing an attention-based recurrent encoder. This architecture integrates proprioceptive and exteroceptive data, allowing the robot to anticipate terrain characteristics and dynamically adapt its movement. Such an approach foregoes manual heuristics, enabling more seamless data combination. The training process employs privileged learning—initially providing the teacher policy with comprehensive environmental data, including ground-truth terrain states, followed by student policy training that mimics the teacher without relying on this privileged information. The student policy uses a learned belief state to compensate for noisy, incomplete, or deceptive sensory data.

Results and Experimental Verification

The paper showcases compelling results, validating the controller across multiple environments and conditions. Notably, the controller was deployed in the Swiss Alps, where it completed a 2.2 km hiking trail with a 120 m elevation gain, aligning with human performance metrics. The paper details the robot's robust traversal of diverse terrains, including complex urban settings and subterranean environments, with zero incidences of failure.

Quantitative experiments further affirm the controller’s capabilities. Comparing success rates of surmounting obstacles, the research highlights significant improvements over a proprioceptive baseline. The controller successfully manages step heights up to 30.5 cm and maintains speed and stability across varied obstacle courses, demonstrating both predictive and corrective adjustments to terrain changes.

Implications and Future Directions

This research has profound implications for the deployment of legged robots in unpredictable environments, enhancing operational range and application versatility—from exploration and rescue missions to industrial inspections. The integration of exteroceptive and proprioceptive sensing surpasses previous limitations, offering pathways to terrain-agnostic, fast locomotion strategies.

Future work could explore deeper exploitation of sensory input, possibly through end-to-end learning of raw sensor data and enhancing uncertainty management. By refining perception layers and incorporating terrain-specific adaptations, potential exists to improve capabilities in navigating even more extreme environments.

In conclusion, this paper contributes significantly to the field by demonstrating a practical, flexible locomotion framework that excels in natural and engineered settings. As robotic systems continue to evolve, such integrative approaches hold promise for expanding the autonomy and functionality of robotic agents in the field.