Four-legged robot moves chairs with contact-seeking training, a post reports

The method guides the robot toward possible contact points, then gradually reduces that guidance during training so it can focus on completing the task, according to the post.

1 Source, 43d ago, first seen 43d ago

TLDR

The post describes a reinforcement-learning problem: when task rewards stay at zero until contact, a robot can get stuck optimizing smoothness and energy penalties instead. It credits researchers from the University of Pisa, ETH Zürich and Nvidia with adding a separate “critic,” or evaluator, trained on contact-seeking rewards. Its influence decreases during training, shifting emphasis toward task performance. According to the post, candidate contact points come from a general-purpose grasping algorithm. It reports tests on a real four-legged robot with a manipulator that moved chairs, transferred to unseen IKEA furniture without additional training, and recovered from failed contact attempts.

Today's Rank

—

Not ranked yet

Combined views

—

1 Source, first seen 43d ago

1 Source

@stepjamUKTeaching a quadruped robot to push an object it can't grasp is a hard RL exploration problem. The task reward stays at zero until contact happens, so a standard single-critic PPO ends up optimizing smoothness and energy penalties instead, and never finds contact at all. Members from @Unipisa, ETH Zürich & @nvidia address this with a separate exploration critic trained on a dense contact-seeking reward, guiding the end-effector toward candidate contact points, then decaying its weight over training so the policy shifts from contact-seeking to task-optimal once the physics has actually been found. Candidate points come from a general-purpose grasping algorithm, so the approach generalizes across object geometries without hand-tuning per task. Validated on a real quadrupedal mobile manipulator transporting chairs, transferring zero-shot to unseen IKEA furniture, recovering from failed contact attempts, and remaining stable under loads beyond the robot's rated capacity. A clean approach to a common problem: rather than hand-shaping one messy scalar reward, split exploration from task performance and let one decay into the other. Paper: https://tolomeis.github.io/contact-guided-exp/assets/contact_guided_exp_RAL.pdf Project page: https://tolomeis.github.io/contact-guided-exp/43d

Today's Rank

—

Not ranked yet

Combined views

—

1 Source, first seen 43d ago

— likes

— comments

— saves

— reposts

— comments

— saves

— reposts

Sentiment

Positive——Negative

Summary

Not enough discussion yet.

No sentiment analysis available yet.

Sentiment

Positive——Negative

Summary

Not enough discussion yet.

No sentiment analysis available yet.

Sentiment

Positive——Negative

Summary

Not enough discussion yet.

No sentiment analysis available yet.

Sentiment

Positive——Negative

Summary

Not enough discussion yet.

No sentiment analysis available yet.

Today's Rank

—

Not ranked yet

Today's Rank

—

Not ranked yet

1 Source

@stepjamUKTeaching a quadruped robot to push an object it can't grasp is a hard RL exploration problem. The task reward stays at zero until contact happens, so a standard single-critic PPO ends up optimizing smoothness and energy penalties instead, and never finds contact at all. Members from @Unipisa, ETH Zürich & @nvidia address this with a separate exploration critic trained on a dense contact-seeking reward, guiding the end-effector toward candidate contact points, then decaying its weight over training so the policy shifts from contact-seeking to task-optimal once the physics has actually been found. Candidate points come from a general-purpose grasping algorithm, so the approach generalizes across object geometries without hand-tuning per task. Validated on a real quadrupedal mobile manipulator transporting chairs, transferring zero-shot to unseen IKEA furniture, recovering from failed contact attempts, and remaining stable under loads beyond the robot's rated capacity. A clean approach to a common problem: rather than hand-shaping one messy scalar reward, split exploration from task performance and let one decay into the other. Paper: https://tolomeis.github.io/contact-guided-exp/assets/contact_guided_exp_RAL.pdf Project page: https://tolomeis.github.io/contact-guided-exp/43d