Reproduce “LocoTouch: Learning Dynamic Quadrupedal Transport with Tactile Sensing” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Changyi Lin, Yuxin Ray Song, Boda Huo, Mingyang Yu, Yikai Wang, Shiqi Liu, Yuxiang Yang, Wenhao Yu and 4 more
Researchers from Carnegie Mellon University and Google DeepMind developed LocoTouch, a system enabling quadrupedal robots to dynamically transport diverse unsecured objects using a custom high-density tactile sensor and a reinforcement learning pipeline. It achieved robust zero-shot sim-to-real transfer for varied objects across uneven terrains and sustained locomotion, with the tactile student policy reaching up to 96.7% success in challenging transport tasks.
Key insights
- High-density, distributed tactile sensing on the robot's back is crucial for perceiving and adapting to the dynamic state of unsecured objects during transport.
- Efficiently simulating contact spread using an "expanded collision model" is vital for effective tactile-aware policy learning in reinforcement learning, bridging the sim-to-real gap.
- An adaptive gait reward function that promotes synchronized diagonal and alternating lateral leg contacts, informed by an object-task-aware symmetricity score, improves gait symmetry and task performance for dynamic object transport.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.