HORA

A reinforcement-learning method that uses only fingertip and joint proprioception to keep rotating a variety of objects in-hand.

HORA was released in October 2022 by Haozhi Qi, Jitendra Malik, and colleagues at UC Berkeley and Meta AI, published at CoRL 2022. The goal is to have the Allegro four-fingered dexterous hand rotate an object continuously about one axis using only its fingertips. The method follows Rapid Motor Adaptation (RMA): first, a base policy is trained with reinforcement learning in simulation, with access to privileged information such as the object's size, mass, and friction; then an adaptation module is trained to infer those same object properties using only a recent history of joint proprioception. The policy is trained in simulation on cylinders alone, yet deploys to the real robot with no fine-tuning, rotating objects that vary in size, shape, and weight; a stable “finger gait” emerges naturally during training. It is a representative example of sim-to-real transfer for in-hand manipulation.

ExampleA policy that only ever saw cylinders in simulation, once deployed on a real Allegro hand, can rotate everyday objects of varying shapes and weights without using a camera at all.

Also called

In-Hand Object Rotation via Rapid Motor Adaptation

Related

Rapid Motor Adaptation · In-hand Manipulation · Dexterous Manipulation · Privileged Information · Sim-to-Real Transfer · Allegro Hand

Sources

In-Hand Object Rotation via Rapid Motor Adaptation (arXiv 2210.04887)
HORA project page

As of

2022-10