Bimanual manipulation

Two limbs create one manipulation system#

Some tasks need more than two independent single-arm motions. One hand may stabilize a container while the other opens its lid. Both hands may carry opposite ends of a flexible object, or bring two parts together for assembly. In each case, the state and action of one limb change what the other limb can safely do.

Bimanual coordination can be symmetric, as when both hands lift a box, or asymmetric, as when one hand holds and the other inserts. It may also include role changes during a task. The defining feature is coordinated robotic manipulation, not simply the presence of two robot arms.

Coordination can come from control or demonstration#

A conventional controller can impose geometric, force, and collision constraints on both arms. A learned policy can instead infer coordinated actions from demonstrations. The ALOHA study introduced a low-cost teleoperation system and Action Chunking with Transformers for fine-grained two-arm tasks such as battery insertion and cable routing.

Benchmarks also isolate the extra coordination problem. PerAct2 extends a simulated manipulation benchmark with tasks that require spatial and temporal coordination between two arms. Its reported results concern those benchmark tasks and policy implementation, not all bimanual robots.

Two arms add more than twice the complexity#

The combined action space is larger, and the arms can collide with each other. Timing errors can stretch, tear, drop, or jam a shared object. Calibration errors between the two arm frames also matter when both end effectors constrain the same rigid part.

Training data must capture cooperation as well as each arm's motion. Teleoperation can provide paired demonstrations, but the operator interface and viewpoint shape the resulting data. A successful bimanual policy still depends on sensing, gripper design, control frequency, object properties, and the distribution of tasks represented during training.