Quadrupedal Footstep Planning Using Learned Motion Models of a Black-Box Controller

Ilyass Taouil, Giulio Turrisi, Daniel Schleich, Victor Barasuol, Claudio Semini, Sven Behnke

Abstract

High-performance quadrupedal robots often utilize proprietary, "black-box" locomotion controllers that accept high-level velocity commands but lack interfaces for precise footstep placement. This limitation hinders their ability to navigate complex terrains such as gaps or stepping stones. This paper presents a methodology to bridge this gap by learning the motion behavior of a black-box controller using Multivariate Linear Regression (MLR). By predicting footstep displacements and Center of Mass (CoM) trajectories from velocity commands, we enable an $A^*$ search-based planner to navigate irregular terrain. Experimental results on a Unitree Aliengo robot demonstrate high prediction accuracy (2cm RMSE) and a 100% success rate in crossing challenging obstacles with real-time planning performance.

Introduction

Legged robots like the Unitree Aliengo and Boston Dynamics Spot have demonstrated remarkable agility and stability. However, many of these platforms are shipped with pre-installed, proprietary controllers that treat locomotion as a reactive, blind process. While these controllers excel at maintaining balance on flat or uneven ground via proprioception, they offer limited control authority to the end-user. Specifically, users can typically only command body velocities ($v_x, v_y, \omega$), rather than specific footfall locations.

This lack of precise placement is a significant hurdle for "perceptive" locomotion—tasks where the robot must avoid specific hazards like gaps, cracks, or small steps. In "Quadrupedal Footstep Planning using Learned Motion Models of a Black-Box Controller," Taouil et al. (IROS 2023) propose a system identification approach that models the robot's response to velocity commands, allowing for deliberate footstep planning without requiring access to the underlying control code.

Methodology

The proposed pipeline consists of three integrated components: motion model learning, perception-based mapping, and heuristic search planning.

1. Learning the Motion Models

To predict how the robot will move in response to specific commands, the researchers developed three distinct models using Multivariate Linear Regression (MLR):

  • Footstep Displacement Model (FDM): Predicts the $(x, y)$ coordinates of future foot landings.
  • CoM Displacement Model (CDM): Predicts the change in the robot's Center of Mass position.
  • CoM Velocity Model (CVM): Predicts the resulting body velocity.

A critical insight in this modeling phase was the inclusion of "state history." Because physical controllers exhibit latency and acceleration ramps, the models take the previous velocity command ($t-1$) as an input alongside the current command. This allows the linear model to capture the non-ideal dynamics of the black-box system. Surprisingly, the authors found that simple MLR outperformed complex non-linear models, suggesting that the relationship between velocity commands and physical displacement in these controllers is highly consistent.

2. Perception and Cost Mapping

The robot utilizes an onboard depth camera to generate a 2.5D elevation map of its environment. This map is processed into a "foot costmap," where terrain features are evaluated for safety. Areas representing gaps, steep inclines, or edges are assigned high cost or flagged as untraversable, ensuring the planner avoids placing feet in unstable locations.

3. $A^*$ Planning and Execution

The system employs a variant of the $A^*$ search algorithm to find an optimal sequence of velocity commands. The planner explores a tree of possible commands, using the learned motion models to "roll out" the predicted trajectory and footfall locations. The search objective is to reach a goal state while ensuring all predicted footsteps land in low-cost zones on the map. The system operates on a receding horizon: it plans several steps into the future but only executes the immediate next command before re-planning, which provides robustness against model inaccuracies and sensor noise.

Experimental Results

The framework was evaluated using a Unitree Aliengo quadruped in three distinct indoor scenarios involving symmetrical and asymmetrical gaps and steps.

  • Accuracy: The learned MLR models achieved a Root Mean Squared Error (RMSE) of approximately 2cm for footstep predictions, providing the precision necessary for narrow stepping targets.
  • Reliability: The robot achieved a 100% success rate across all test scenarios. It successfully navigated obstacles that would have caused the default "blind" controller to fail.
  • Efficiency: The planning process is computationally efficient, with an average computation time of less than 2ms for a 7-step horizon. This low latency allows the robot to maintain fluid motion while reacting to the environment in real-time.

Discussion and Limitations

While the approach effectively enables perceptive locomotion for black-box systems, it is subject to the constraints of the underlying controller. Since the planner can only send velocity commands, it cannot modify the robot's gait pattern (e.g., switching from a trot to a crawl) or adjust step height and body posture. This limits the robot's ability to navigate very high obstacles or low-clearance environments.

Furthermore, the current safety margins are static. Future work may involve the use of Bayesian regression to estimate model uncertainty dynamically. This would allow the robot to adopt more conservative behaviors when the motion model's predictions are less certain, further increasing reliability in novel environments.

Conclusion

By treating a proprietary quadruped controller as a system to be identified rather than a barrier, this research provides a practical path forward for deploying commercial robots in complex, unstructured environments. The combination of simple, effective linear modeling and robust heuristic search proves that high-level planning is possible even with limited control authority.

Summary updated

Abstract

Legged robots are increasingly entering new domains and applications, including search and rescue, inspection, and logistics. However, for such a systems to be valuable in real-world scenarios, they must be able to autonomously and robustly navigate irregular terrains. In many cases, robots that are sold on the market do not provide such abilities, being able to perform only blind locomotion. Furthermore, their controller cannot be easily modified by the end-user, requiring a new and time-consuming control synthesis. In this work, we present a fast local motion planning pipeline that extends the capabilities of a black-box walking controller that is only able to track high-level reference velocities. More precisely, we learn a set of motion models for such a controller that maps high-level velocity commands to Center of Mass (CoM) and footstep motions. We then integrate these models with a variant of the <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$A$</tex> * algorithm to plan the CoM trajectory, footstep sequences, and corresponding high-level velocity commands based on visual information, allowing the quadruped to safely traverse irregular terrains at demand.
BibTeX
@inproceedings{taouil2023quadrupedal,
  author = {Ilyass Taouil and Giulio Turrisi and Daniel Schleich and Victor Barasuol and Claudio Semini and Sven Behnke},
  title = {{Quadrupedal Footstep Planning Using Learned Motion Models of a Black-Box Controller}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2023},
  url = {https://openreview.net/forum?id=xGWNwn0gqbk},
}
Figure