Learning Tactile-Aware Quadrupedal Loco-Manipulation Policies

A team of researchers from Columbia, CAU, IIT, and elsewhere has introduced a tactile-aware learning pipeline for quadrupedal robots that significantly improves contact-rich manipulation. The approach addresses a fundamental limitation of vision and proprioception: they cannot capture the subtle, evolving forces during object interaction. By integrating tactile sensing directly into the policy, the robot learns not only what to do but how the contact should feel over time.

The system has two stages. First, a high-level visuotactile policy is trained from real human demonstrations; it predicts end-effector trajectories and tactile interaction cues. Second, a whole-body control policy is trained via large-scale RL in simulation, learning to track those trajectories and touch patterns. This policy transfers zero-shot to real hardware. Tested on tasks like in-hand reorientation with insertion, valve tightening, and delicate object manipulation, the tactile-aware policy achieved a 28.54% average improvement over vision-only and visuotactile baselines. The work suggests that touch is a critical missing modality for dexterous robot manipulation.