State Estimation Transformers for Agile Legged Locomotion

TL;DR

This paper introduces State Estimation Transformers (SET), a novel Transformer-based approach for accurately predicting complex robot states during dynamic quadruped locomotion, enabling advanced skills like jumping.

Contribution

The paper presents a new Transformer-based architecture for state estimation that improves accuracy and transferability in highly dynamic quadruped robot tasks.

Findings

SET outperforms existing methods in estimation accuracy.

SET achieves higher success rates in real-world jumping tasks.

Transformer-based estimation enhances dynamic locomotion capabilities.

Abstract

We propose a state estimation method that can accurately predict the robot's privileged states to push the limits of quadruped robots in executing advanced skills such as jumping in the wild. In particular, we present the State Estimation Transformers (SET), an architecture that casts the state estimation problem as conditional sequence modeling. SET outputs the robot states that are hard to obtain directly in the real world, such as the body height and velocities, by leveraging a causally masked Transformer. By conditioning an autoregressive model on the robot's past states, our SET model can predict these privileged observations accurately even in highly dynamic locomotions. We evaluate our methods on three tasks -- running jumping, running backflipping, and running sideslipping -- on a low-cost quadruped robot, Cyberdog2. Results show that SET can outperform other methods in estimation accuracy and transferability in the simulation as well as success rates of jumping and triggering a recovery controller in the real world, suggesting the superiority of such a Transformer-based explicit state estimator in highly dynamic locomotion tasks.