ProNav: Proprioceptive Traversability Estimation for Legged Robot Navigation in Outdoor Environments

TL;DR

This work presents ProNav, a proprioception-driven traversability estimator for outdoor legged robots that uses joint positions, knee forces, and battery current to assess terrain stability and motion resistance. By projecting a compact 9-dimensional proprioceptive feature vector through PCA to a 2D space, it identifies low-variance regions and ellipses corresponding to stable gait-Terrain pairs, enabling adaptive gait selection to enhance stability and reduce energy while predicting imminent crashes. The approach is designed to complement exteroceptive planners, yielding a resilient exo-proprioceptive navigation framework that improves success rates (up to 40% in vegetation-rich environments) and reduces energy consumption (up to 15.1%) relative to exteroceptive baselines. Extensive experiments on a Spot platform across granular, rocky, and vegetated terrains demonstrate robust performance, supported by ablations that underscore the importance of current and hip-position signals for traversability estimation and crash avoidance.

Abstract

We propose a novel method, ProNav, which uses proprioceptive signals for traversability estimation in challenging outdoor terrains for autonomous legged robot navigation. Our approach uses sensor data from a legged robot's joint encoders, force, and current sensors to measure the joint positions, forces, and current consumption respectively to accurately assess a terrain's stability, resistance to the robot's motion, risk of entrapment, and crash. Based on these factors, we compute the appropriate robot gait to maximize stability, which leads to reduced energy consumption. Our approach can also be used to predict imminent crashes in challenging terrains and execute behaviors to preemptively avoid them. We integrate ProNav with an exteroceptive-based method to navigate real-world environments with dense vegetation, high granularity, negative obstacles, etc. Our method shows an improvement up to 40% in terms of success rate and up to 15.1% reduction in terms of energy consumption compared to exteroceptive-based methods.

Figures (9)

• Figure 1: Comparison of our method ProNav with other methods navigating a Spot robot through dense vegetation: ProNav adapts between two gaits: trot (in red), and amble (in green), Spot's in-built planner (black), GA-Nav guan2022ga (trot: yellow, crawl: brown), RFCkertesz2016rigidity (trot: light blue and amble: dark blue), and VERN sathyamoorthy2023vern (purple). In this scenario, we observe that our method successfully traverses the dense vegetation due to its efficient gait adaptation and accurate proprioception-based traversability estimation.
• Figure 2: Images (a)-(c) depict the RGB images captured sequentially from the robot's camera. (d) Plot of the fluctuations in knee force readings experienced by the robot while traversing the terrain. The high fluctuations represent instances when the robot became unstable. This shows that visually identical terrains could have different stability characteristics.
• Figure 3: (a) Changes in the hip X-axis position of the robot while traversing grass (green box), and rocks (brown box) plotted over time. (b) Force exerted by the four knee actuators while traversing grass (green box), and rocks (brown box). Steady readings observed on stable grass terrain reflect ease of traversal, while the increased volatility and noticeable spikes on the rocky terrain are indicative of increased resistance and slippage, causing variable load on the actuators. (c) Changes in the hip Y-axis position of the robot while traversing dense vegetation (violet box), and concrete (gray box). High fluctuations are observed while traversing dense vegetation due to the legs' entanglement instances. Conversely, a steady and consistent reading is observed during concrete traversal.
• Figure 4: The average current consumption in amperes, with 95% confidence interval as the robot traverses concrete, grass, and sand. Lower current consumption on concrete indicates ease of traversal. However, higher values on sand highlight increased resistance and energy usage. Additionally, a consistent trend in current consumption is exhibited while using crawl, trot, and amble on various terrains.
• Figure 5: (a) PCA applied to key proprioceptive metrics (hip actuator positions, knee actuator force, and battery current) across two different terrains when using the trot gait. The variances along the two principal components indicate the level of stability on a terrain. (b) The figure shows the shift in the PCA distribution between stable navigation (grey points), before a crash (yellow), which represents 3 seconds before the crash, and 10 seconds after a crash (red), where a robot falls to the ground. If the robot's proprioceptive signals lie outside the ellipse $\Gamma_{safe}$, the robot is heading towards a crash.