Stabilize to Act: Learning to Coordinate for Bimanual Manipulation

Abstract

The ability to coordinate meaningfully between two hands unlocks a variety of skills to humans. Bimanual robotics opens the door to accomplishing a similarly large set of manipulation tasks. However, constructing control policies for dual arm autonomous systems brings inherent difficulties---the high-dimensionality of the bimanual action space adds complexity to both model-based and data-driven methods. We counteract this challenge by drawing inspiration from humans to propose a novel role assignment framework: a stabilizing arm holds an object in place to simplify the environment while an acting arm executes the task. We instantiate this framework with BimanUal Dexterity from Stabilization (BUDS), which uses a learned restabilizing classifier to alternate between updating a learned stabilization position to keep the environment unchanged, and accomplishing the task with an acting policy learned from demonstrations.

We evaluate BUDS on four dexterous bimanual tasks on real-world robots and present experiments highlighting overall task success and policy generalizability across objects within a class. Given 20 demonstrations, BUDS achieves 76.9% success on completing a wide variety of physical bimanual tasks, and achieves 52.7% success on when generalizing to out-of-distribution objects. BUDS is 56.0% more successful than a unstructured baseline that instead learns a BC stabilizing policy due to the precision required of these complex tasks.

BUDS learns to coordinate by assigning roles to each arm: a stabilizing arm holds a point stationary for a period of time while an acting arm acts. The stabilizing position to hold stationary is learned as a keypoint on an overhead image, and is instantiated with a ResNet architecture. Then, a noncompliant controller holds this point stable while an acting arm rolls out a policy learned from 20 expert single-arm demonstrations with a BC-RNN architecture. These two actions comprise a bimanual action at a single time step. Finally, we inject flexbility into the stabilizing policy with a Restabilizing Classifier, which determines via visual feedback at every time step when a stabilizing position is no longer effective and a new point should be detected. This final piece allows BUDS to tackle more challenging tasks, such as making multiple cuts moving back up the length of a vegetable, as shown here.

Experiments

We present experiments on four complex and diverse bimanual tasks that all require dynamic stabilizing policies and high-precision acting policies. Together, this set represents a wide variety of bimanual tasks. In these experiments, we visualize the task performance on in-distribution objects that are seen during training. Later, we stress test BUDS's generalizability with more out of distribution objects (below).

We compare to a BC-Stabilizer baseline, where both arms are each controlled by a policy learned via imitation learning. In other words, this baseline replaces BUDS's keypoint model with a stabilizing policy learned with a BC-RNN architecture.

Generalizing to OOD Objects

We stress test the ability of BUDS to generalize to out-of-distribution (OOD) objects completely unseen during training. BUDS is able to generalize across OOD objects with highly varied visual appearances, materials, geometries, and dynamics.