Reproduce “Helpful DoggyBot: Open-World Object Fetching using Legged Robots and Vision-Language Models” 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

Qi Wu, Zipeng Fu, Xuxin Cheng, Xiaolong Wang, Chelsea Finn

The Helpful DoggyBot system enables a quadrupedal robot to perform open-world object fetching tasks in diverse indoor environments by integrating a custom gripper, a learned whole-body controller for agile locomotion, and pre-trained Vision-Language Models (VLMs) for semantic understanding and command generation. This approach allows the robot to generalize zero-shot to unseen environments and objects, demonstrating 60% success rates on complex real-world tasks like climbing a bed to retrieve an item.

Key insights

  • A simple, lightweight gripper design can effectively enable object manipulation on a quadruped without sacrificing its agility or introducing excessive complexity.
  • A two-phase reinforcement learning approach with policy distillation from privileged information to egocentric depth significantly improves a quadruped's whole-body control, enhancing its traversability and reachability in complex indoor environments.
  • Leveraging off-the-shelf pre-trained Vision-Language Models allows for robust zero-shot generalization in open-vocabulary object perception and task execution, bridging the sim-to-real gap for semantic understanding.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.