Robot dog learns to hook and drag IKEA chairs, succeeding in 40/58 trials

ChongZzZhang · x · 2026-09-02

Core Achievement: The ALMA quadruped robot learned to transport non-prehensile objects like IKEA chairs using a contact-guided exploration strategy. It succeeded in 40 out of 58 real-world trials on 4 unseen objects.
Technical Method: Uses a Multi-Critic Reinforcement Learning (RL) framework. A dedicated exploration critic rewards finding contact early in training. Once interaction points are reliably discovered, this objective fades, and the policy focuses on moving the object to the goal.
Behavior: The robot uses its arm to discover useful contact points, then pushes, pulls, and hooks bulky objects while coordinating its whole body. It can even reposition and retry with a larger swing if a hook slips.
Use Cases: Non-prehensile manipulation is versatile for heavy or bulky objects. The paper also demonstrates box pushing and dishwasher opening tasks.

Paper Link

More from Embodied