LocoTouch: Learning Dexterous Quadrupedal Transport with Tactile Sensing

Carnegie Mellon University University of Washington Google Deepmind

Abstract

Quadrupedal robots have demonstrated remarkable agility and robustness in traversing complex terrains. However, they remain limited in performing object interactions that require sustained contact. In this work, we present LocoTouch, a system that equips quadrupedal robots with tactile sensing to address a challenging task in this category: long-distance transport of unsecured cylindrical objects, which typically requires custom mounting mechanisms to maintain stability.

For efficient large-area tactile sensing, we design a high-density distributed tactile sensor array that covers the entire back of the robot. To effectively leverage tactile feedback for locomotion control, we develop a simulation environment with high-fidelity tactile signals, and train tactile-aware transport policies using a two-stage learning pipeline. Furthermore, we design a novel reward function to promote stable, symmetric, and frequency-adaptive locomotion gaits.

After training in simulation, LocoTouch transfers zero-shot to the real world, reliably balancing and transporting a wide range of unsecured, cylindrical everyday objects with broadly varying sizes and weights. Thanks to the responsiveness of the tactile sensor and the adaptive gait reward, LocoTouch can robustly balance objects with slippery surfaces over long distances, or even under severe external perturbations.

Video Overview

Agile Transport in Cluttered Environments

LocoTouch achieves to transport a 1.4kg unsecured cylinder through dense obstacles with sharp turns.

Long-Distance Transport of Unsecured Objects

Transport a 1.4kg cylinder over 40m.

Transport a sleppery bottle over 60m.

Policy Generalization

Transport a wide range of cylindrical objects with different sizes and weights.
Diameters: 0.03-0.18m; Lengths: 0.10-1.26m; Weights: 0.03-1.45kg.

Policy Robustness

Symmetric, Frequency-Adaptive Gaits

Teacher Policies Trained with and without Symmetricity Function

We deploy both teacher policies on real robots with a forward velocity command of 0.3m/s.
The baseline policy exhibits lateral drifting due to asymmetric deformation of the foot pairs.
In contrast, our policy with symmetric gaits tracks the forward velocity command accurately.
Details about the adaptive gait reward and symmetricity function can be found in the paper.

Baseline: Locomotion-Only Policy

The locomotion policy trained without object interaction totally fails to transport objects.

High-Fidelity Tactile Simulation

InterSect: an intersection-only model that detects contact only at cable intersections;
Filtered: a Gaussian-smoothed model that applies a Gaussian filter to the output of InterSect;
Ours: the proposed expanded collision model that enlarges each taxel's effective sensing area.
Our efficient and scalable modeling method generates the most high-fidelity tactile signals.

BibTeX


      @article{lin2025locotouch,
      title={LocoTouch: Learning Dexterous Quadrupedal Transport with Tactile Sensing},
      author={Lin, Changyi and Song, Yuxin Ray and Huo, Boda and Yu, Mingyang and Wang, Yikai and Liu, Shiqi and Yang, Yuxiang and Yu, Wenhao and Zhang, Tingnan and Tan, Jie and Luo, Yiyue and Zhao, Ding},
      journal={arXiv preprint arXiv:2505.23175},
      year={2025}
      }