LocoTouch: Learning Dynamic Quadrupedal Transport with Tactile Sensing
Published in Conference on Robot Learning (CoRL ’25), PMLR 305: 2779–2801, 2025
Paper: PMLR proceedings
Project page: linchangyi1.github.io/LocoTouch
Summary
LocoTouch studies tactile-aware learning for dynamic quadrupedal transport. The core idea is to use distributed tactile sensing to improve robustness during long-horizon carrying tasks, especially under contact uncertainty.
My Contribution
I contributed to both the hardware and learning / sim-to-real pipeline that enable tactile-aware transport:
- Designed and fabricated the distributed tactile hardware (221-taxel piezoresistive array) to provide wide-area, high-density contact feedback over the robot’s back. I iterated on the physical build and wiring/scanning approach to support reliable, repeatable sensing during dynamic motion.
- Improved manufacturability and robustness of the sensing stack, focusing on practical issues that matter for long experiments—repeatability, yield, and stability across runs. In parallel, I supported updates to the PCB and sensing electronics with the goal of reducing cross-talk and increasing frame rate.
- Helped develop the tactile-aware transport policy, contributing to a teacher–student training pipeline (PPO → DAgger) and emphasizing reliable tracking behavior with PD control so the learned policy remains stable during real hardware execution.
Recommended citation: Lin, C.; Song, Y. R.; Huo, B.; Yu, M.; Wang, Y.; Liu, S.; Yang, Y.; Yu, W.; Zhang, T.; Tan, J.; Luo, Y.; Zhao, D. (2025). "LocoTouch: Learning Dynamic Quadrupedal Transport with Tactile Sensing." Conference on Robot Learning (CoRL 2025), PMLR 305: 2779–2801.
Download Paper