Policy learns to find articulations in objects for better robot manipulation
Learning to Explore Hidden Kinematics for Articulated Object Manipulation
Robotics
Summary
Figuring out how objects with moving parts work, like doors or drawers, is hard just by looking since their joints aren’t always obvious. The authors show a way for robots to learn how to explore these hidden joints by interacting with the object and updating their guess based on the motion they see. They represent their understanding as a special motion map and train a robot policy to pick actions that reduce uncertainty about the object’s joints over time. This learned approach works better than previous methods on tests involving doors and drawers, including tricky ones where appearances can be misleading.
What this means in practice
- •For robotics engineers: Enable robots to actively interact with unknown articulated objects to figure out their joint types and parameters for improved manipulation.
- •For automation system developers: Design automated systems that can learn to explore and manipulate objects with hidden or ambiguous joints without detailed prior models.
Authors
Ruiyao Liu, Boshu Lei, Zhuoyang Pan, Kostas Daniilidis
Abstract
The kinematics of an articulated object is often ambiguous from vision alone. Interaction resolves the ambiguity, and active perception methods exploit this by searching for the single action that most sharpens a belief over the kinematic parameters at each step. Such greedy search cannot be extended over a horizon without forward models of the contact and inertial dynamics, which are themselves unknown. We instead amortize action selection into training. We maintain a belief distribution over joint type and parameters, initialized from a generative prior and updated by Bayesian filtering on the observed part motion. To condition the policy on this belief, we render it as a per-point articulation flow field, the motion that the current posterior predicts for every point on the object. Carrying the inductive bias of articulated motion, this representation generalizes better than a latent encoding of the belief or flow tracked from observation. We train the policy with reinforcement learning, rewarding the entropy that each interaction removes from the posterior, so that informative exploration becomes learned behavior rather than a search at every step. Our method outperforms previous approaches across door and drawer manipulation on the PartManip benchmark, and reaches 61.7% success on ArticuRiddle, a new dataset of objects whose appearance implies the wrong articulation, against 44.4% for the best previous method. Project Website: https://hiddenkinematics.github.io/