LocoTouch: Learning Dexterous Quadrupedal Transport with Tactile Sensing (2505.23175v1)

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.

• The paper presents LocoTouch, a system that enables quadruped robots to transport unsecured objects by integrating high-density tactile sensing and adaptive reinforcement learning policies trained in simulation and transferred zero-shot to the real world.
• A key innovation is a novel simulation model that captures tactile signal coupling and an adaptive gait reward mechanism allowing flexible trotting behavior for dynamic object transport.
• LocoTouch successfully transports varied objects over long distances on challenging terrain, outperforming locomotion-only and tactile-deprived systems and showing robustness to perturbations, validating the importance of tactile feedback for dexterous tasks via zero-shot sim-to-real transfer.

Overview of "LocoTouch: Learning Dexterous Quadrupedal Transport with Tactile Sensing"

"LocoTouch: Learning Dexterous Quadrupedal Transport with Tactile Sensing" presents a sophisticated system for quadrupedal robots designed to enhance their ability to transport a variety of unsecured cylindrical objects over long distances, leveraging tactile sensing capabilities. Despite advancements in quadrupedal locomotion that enable navigation across complex terrains, object manipulation requiring sustained contact remains a significant challenge. This paper addresses this gap by integrating tactile sensing into the control and perception strategies of quadrupedal robots, advancing their ability to perform dexterous transport tasks.

Key Components

Tactile Sensor Design:

The researchers have developed a high-density tactile sensor array that extensively covers the robot's back, comprising 221 taxels that provide detailed contact feedback. This sensor design incorporates piezoresistive film technology for effective force detection, harnessing voltage variations at intersections of conductive layers.

Simulation and Policy Development:

A novel simulation model captures the tactile signal coupling seen in real-world sensing, optimizing the realism of contact feedback without the need for computationally prohibitive processing techniques. The system employs a two-stage learning pipeline—initializing tactile policy training with a teacher-student approach in a simulated environment, followed by leveraging adaptive reinforcement learning (RL) strategies to transfer the learned policy to real-world implementations efficiently.

Adaptive Gait Reward:

A significant contribution of this work is the adaptive gait reward mechanism that facilitates symmetric trotting behavior without the constraints of predefined patterns. This is critical for ensuring agility and responsiveness in dynamic environments where transport policies must adapt to variable object and terrain conditions.

Results and Implications

The research demonstrates successful zero-shot transfer of dexterous transport policies from simulation to real-world applications. LocoTouch exhibited robustness in transporting cylindrical objects of various sizes and weights, consistently maintaining balance even under severe perturbations. The tactile feedback allowed for adaptive responses to changes in object dynamics—key to sustaining transport over extensive trajectories and challenging terrains.

Quantitatively, the tactile-enabled policies greatly surpassed both locomotion-only baselines and variants deprived of tactile input, underscoring the necessity of real-time, high-fidelity tactile information for complex task execution. The system's responsiveness was further exhibited during rigorous manipulation challenges, such as navigating obstacles and adapting to sharp velocity command shifts.

Future Directions

The implications of this work extend toward broader applications in unstructured environments, suggesting pathways for integrating multi-modal inputs, such as vision alongside tactile sensing, to enable more comprehensive interactions. Furthermore, extending dense tactile coverage to additional areas on the robots presents exciting prospects for whole-body manipulation and omnidirectional tactile interaction, potentially transforming the capabilities and use cases for quadrupedal robotic systems in commercial and research domains.

Overall, LocoTouch represents a significant step forward in tactile sensing technology and robotics control strategies, providing both theoretical and practical contributions to autonomous robotic systems capable of interacting with their environments beyond mere locomotion.