humanoidsdata.com

Search

Search datasets, articles, and glossary terms for humanoids and embodied AI.

← All glossary terms

Hardware & control

Bimanual manipulation

Bimanual manipulation is the use of two hands or robot manipulators to carry out one manipulation task with spatial, temporal, or force coordination between them. The two sides may play symmetric roles, such as lifting a large object together, or asymmetric roles, such as one hand holding an object while the other operates a tool.

Also known as: dual-arm manipulation, dual arm manipulation, two-arm manipulation

Updated

Two arms do not automatically make one bimanual task

The literature does not use one perfectly settled boundary. The dual-arm manipulation survey distinguishes independent actions from coordinated actions and notes that some narrower definitions reserve “bimanual” for two manipulators physically interacting with the same object.

The practical distinction is coordination. Two arms sorting unrelated objects at the same time can be treated as parallel single-arm tasks. Folding a garment, opening a container, passing an object between hands or stabilising a part during insertion requires the two action streams to satisfy shared timing, geometry or contact constraints.

Coordination can be loose or tightly coupled

The bimanual manipulation taxonomy separates uncoordinated actions from coordinated ones. Loose coupling may require the hands to meet at a handover point or reach a state at the same time. Tight coupling adds continuing trajectory and force dependencies, as when both hands carry one rigid object.

Roles also vary. Symmetric actions give both hands similar jobs. In asymmetric actions, one hand can establish a reference frame or stabilise the object while the other performs a more precise movement. A policy and evaluation protocol should preserve that distinction rather than treating the arms as interchangeable.

Bimanual data needs a shared timeline

The ALOHA project demonstrates how bimanual learning depends on synchronised observations and commands from both sides during precise, contact-rich tasks. A dataset should record both arm and hand states, both action streams, camera views, object state, contacts, task phase and outcome on one clock.

Independent normalisation or dropped samples can alter the relative timing that made a demonstration work. For humanoids, torso motion, balance and whole-body constraints may also be part of the task even when the visible interaction happens at the hands.

Sources