AsymDex: Leveraging Asymmetry and Relative Motion in
Learning Bimanual Dexterity
Anonymous authors
Abstract
We present Asymmetric Dexterity (AsymDex), a novel reinforcement learning (RL) framework that can efficiently learn asymmetric bimanual skills for multi-fingered hands without relying on demonstrations, which can be cumbersome to collect. Two crucial ingredients enable AsymDex to reduce the observation and action space dimensions and improve sample efficiency. First, AsymDex leverages the natural asymmetry found in human bimanual manipulation and assigns specific and inter-dependent roles to each hand: a facilitating hand that moves and reorients the object, and a dominant hand that performs complex manipulations on said object. Second, AsymDex defines and operates over relative observation and action spaces, facilitating responsive coordination between the two hands. Further, AsymDex can be easily integrated with recent advances in grasp learning to handle both the object acquisition phase and the interaction phase of bimanual dexterity. Unlike existing RL-based methods for bimanual dexterity, AsymDex is not tailored to a specific task. Detailed experiments on four simulated asymmetric bimanual dexterous manipulation tasks reveal that AsymDex consistently outperforms strong baselines that challenge its design choices, in terms of success rate and sample efficiency.
Video
AsymDex_video.mp4Method
AsymDex has two crucial ingredients, asymmetry and relative pose:
Asymmetry
We introduces asymmetry by defining a dominant hand and a facilitating hand. While the facilitating hand learns to reposition and reorient the object, the dominant hand learns complex manipulation skills. AsymDex only learns to control the hand bases of two hands and the finger of dominant hand, with the assumption that facilitating hand has already grasped the object. Hence AsymDex doesn not need to learn finger motion of facilitating hand.
Relative pose
We introduce relative pose by computing observations and actions in object frame, which we define as relative observation and action spaces. And we use a designed bimanual controller to convert the relative actions into common actions.
Based on asymmetry structure, relative pose further reduce the dimensionality of the observation and action spaces.
We also leverage the observation that, bimanual manipulation in practice is composed of two distinct phases:
The acquisition phase in which objects are grasped from surfaces.
The interaction phase where the two hands coordinate to perform the bimanual task.
Unlike many existing methods that ignore the acquisition phase, we show that this decomposition enables AysmDex to be seamlessly integrated with learned grasping policies to enable fluent execution.
Experiment
We evaluate AsymDex on four complex bimanual dexterous manipulation tasks (block in cup, stack, bottle cap, and switch), and compare the training results with strong baselines that challenge our key design choices.
Block in cup
Stack
Bottle cap
Switch
Learning bimanual coordination
Monolithic
A baseline that utilizes neither relative pose nor asymmetry.
Asym-w/o-rel
A baseline that only utilizes asymmetry without relative pose.
Rel-w/o-Asym
A baseline that only utilizes relative pose without asymmetry.
AsymDex
Our framework, which utilizes both asymmetry and relative pose.
Combining interaction phase with grasping phase
One-stage-monolithic
This baseline uses a single policy to learn both the grasping and interaction phases for both hands
Two-stage-monolithic
This baseline utilizes phase decomposition, but leverages neither asymmetry nor relative motion
Two-stage-AsymDex
Our method utilizes phase decomposition, and leverages both asymmetry and relative motion.
