Separate exploration critic lets quadrupeds learn to push ungraspable objects (Pisa/ETH/NVIDIA)

stepjamUK · x · 2026-09-07

Teaching a quadruped to push an ungraspable object is a hard RL exploration problem: task reward stays zero until contact, so standard single-critic PPO optimizes smoothness/energy penalties instead and never finds contact.
Researchers from Unipisa, ETH Zürich and NVIDIA train a separate exploration critic on a dense contact-seeking reward, guiding the end-effector toward candidate contact points, then decay its weight so the policy shifts to task-optimal once contact physics is found.
Candidate points come from a general-purpose grasping algorithm, so the approach generalizes across object geometries without per-task hand-tuning.
Validated on a real quadrupedal mobile manipulator transporting chairs, transferring zero-shot to unseen IKEA furniture and recovering from failed contacts.

More from Embodied