Gaitor: Learning a Unified Representation Across Gaits for Real-World Quadruped Locomotion

TL;DR

To the best of the knowledge, this is the first work learning a unified and interpretable latent space for multiple gaits, resulting in continuous blending between different locomotion modes on a real quadruped robot.

Abstract

The current state-of-the-art in quadruped locomotion is able to produce a variety of complex motions. These methods either rely on switching between a discrete set of skills or learn a distribution across gaits using complex black-box models. Alternatively, we present Gaitor, which learns a disentangled and 2D representation across locomotion gaits. This learnt representation forms a planning space for closed-loop control delivering continuous gait transitions and perceptive terrain traversal. Gaitor's latent space is readily interpretable and we discover that during gait transitions, novel unseen gaits emerge. The latent space is disentangled with respect to footswing heights and lengths. This means that these gait characteristics can be varied independently in the 2D latent representation. Together with a simple terrain encoding and a learnt planner operating in the latent space, Gaitor can take motion commands including desired gait type and swing characteristics all while reacting to uneven terrain. We evaluate Gaitor in both simulation and the real world on the ANYmal C platform. To the best of our knowledge, this is the first work learning a unified and interpretable latent space for multiple gaits, resulting in continuous blending between different locomotion modes on a real quadruped robot. An overview of the methods and results in this paper is found at https://youtu.be/eVFQbRyilCA.

Figures (8)

• Figure 1: Gaitor's structured latent space forms a planning space in which trajectories for trotting on flat ground, climbing terrain with perception, transitioning to crawl and to pace can be achieved on the real robot. The structure of the latent space during these four scenarios are depicted in the second row. The coloured points in the latent space represent the contact state of the robot and a colour code is provided in the bottom row. The black trajectories in the latent space are parameterised by the gait phase $\phi_k$ and radius $R_k$ at timestep $k$. The latent space trajectory adjusts as the robot climbs terrain to change the swing characteristics. As the robot changes gait, the structure of the latent space changes as does the contact schedule via automatically discovered intermediary gaits.
• Figure 2: There are three components to Gaitor. The first is a variational autoencoder (VAE) to encode the robot states to build the latent space and is comprised of an encoder $\psi_{enc}$ and decoder $\psi_{dec}$, see panels (a) and (c). The second is a terrain encoder $\psi_{ter}$ which represents a rudimentary encoding of the ground ahead of the robot. A learnt planner $\psi_{plan}$ creates trajectories in the latent space, see panel (b). The planner's trajectory exploits the latent-space structure to vary the robot's gait characteristics (footswing heights and lengths) in response to terrain. The updated latent-space trajectory is decoded to predict a set of future robot states ${\mathbf{X}}^+_r(k)$ and contact states ${\mathbf{S}}^+(k)$ using the performance predictor $\psi_{PP}$, see panel (c). The future trajectory is sent to a whole-body controller.
• Figure 3: Terrain processing pipeline. The onboard perception module produces a 2.5D height map of the terrain around the robot. This map is sampled at the footfall positions to estimate the heights of the terrain at these locations, see panels (a) and (b). The control-pitch angle $\theta_c$ is defined as the difference in heights between the front and rear footfalls, see panel (c). These values are input to a second-order filter to create the terrain's encoder input ${\mathbf{X}}_G$ (panel (d)).
• Figure 4: The latent space is disentangled and structured such that displacements in the horizontal $z_0$ axis increase the robot's step height, whereas movements in the vertical axis $z_1$ correlate to the robot's swing length.
• Figure 5: Gaitor deployed on the robot climbing a 12.5cm platform. The planner and latent-space structure act in concert to alter the locomotion trajectories in response to the terrain. The first phase is flat ground where the planner produces a circular ellipse. When the robot steps onto the platform, the planner radius increases dramatically along the horizontal axis, but is reduced in the vertical axis compared to flat ground operation. This results in shorter footstep heights, but longer swing distances. Once both front feet are on the terrain, the latent-space trajectory is circular. The latent-space trajectory synchronised with the robot climbing terrain can be found in this https://youtu.be/eVFQbRyilCA.