Steppability-informed Quadrupedal Contact Planning through Deep Visual Search Heuristics

TL;DR

This work addresses the challenge of planning footholds for quadrupeds in unknown environments by predicting steppability directly in image space. It introduces a synthetic-data pipeline based on primitive shapes to train a semantic segmentation model (DeepLabV3+) that labels steppable, passable, and non-passable regions, and then embeds this perception-space information into an interleaved graph search and trajectory optimization foothold planner. The steppability heuristic guides edge costs, enabling proactive obstacle avoidance and online re-planning, with offline results showing accelerated experience accumulation and online results demonstrating reactive, disturbance-aware navigation. The approach holds promise for faster, interpretable perception-informed planning in legged robotics and motivates future sim-to-real integration and comparisons with traditional point-cloud representations.

Abstract

In this work, we introduce a method for predicting environment steppability -- the ability of a legged robot platform to place a foothold at a particular location in the local environment -- in the image space. This novel environment representation captures this critical geometric property of the local terrain while allowing us to exploit the computational benefits of sensing and planning in the image space. We adapt a primitive shapes-based synthetic data generation scheme to create geometrically rich and diverse simulation scenes and extract ground truth semantic information in order to train a steppability model. We then integrate this steppability model into an existing interleaved graph search and trajectory optimization-based footstep planner to demonstrate how this steppability paradigm can inform footstep planning in complex, unknown environments. We analyze the steppability model performance to demonstrate its validity, and we deploy the perception-informed footstep planner both in offline and online settings to experimentally verify planning performance.

Figures (7)

• Figure 1: Visualization of predicted steppability labels overlaid on irregular stepping stones. Green regions are steppable, meaning we can plan footholds at that location, yellow regions are passable, meaning that we can not plan footholds at that location but can plan swing trajectories over it, and red regions are non-passable, meaning that we can not plan footholds at that location and we also can not plan swing trajectories over it. All four images above show the same environment with (a) - (c) providing side views and (d) providing a top view.
• Figure 2: Results of training. (a) Total training loss over the training iterations. (b) Intersection-over-union (IoU) of all classes over the training iterations.
• Figure 3: Example outputs of learned steppability model. Top row: input depth images to the model. Middle row: ground truth steppability labels of the corresponding column's input depth image. Bottom row: model outputs for the corresponding column's input depth image.
• Figure 4: Offline planning results for Section \ref{['sec:exp1']}. On the left side, results of all planning trials run (a) without and (b) with the steppability heuristic active in the graph search. On the right side, graph search and trajectory optimization solve times for each trial are shown.
• Figure 5: Online tracking performance of offline reference trajectory in simulation.