Bimanual manipulation

Bimanual manipulation is robot manipulation that uses two arms or hands in coordination, either doing complementary jobs, like one hand steadying an object while the other works on it, or sharing the load, like two hands lifting a tray. It is essential for humanoid robots and household tasks, and it roughly doubles the action space a policy has to learn.

What is bimanual manipulation?

Many everyday tasks need two hands: opening a jar, folding a towel, putting on a shoe, threading a needle. Bimanual manipulation is the robot version. A dual-arm robot or a humanoid robot coordinates both arms so that each one does part of the task, and the timing between them matters as much as each arm's motion.
Coordination comes in two broad forms. In asymmetric tasks, one arm holds or steadies while the other acts, as in uncapping a pen. In symmetric tasks, both arms do the same job together, as in lifting a tray with two hands.

Key takeaways

  • Bimanual manipulation coordinates two arms or hands on one task.
  • It roughly doubles the action space, and a policy has to learn the timing between arms along with each arm's motion.
  • Low-cost teleoperation rigs such as ALOHA made large bimanual demonstration datasets practical.

How it works

A bimanual robot's action space combines both arms. A pair of 6-DoF arms with parallel grippers produces a 14-dimensional action, and dexterous hands push it much higher: in one bimanual dataset Voxel51 has explored, the two hands account for 44 of 58 joint positions. Learned policies usually predict both arms in one output so that a single model handles the coordination.
Dexterous hands account for most of a bimanual robot's action dimensions.
Most bimanual data comes from teleoperation. The ALOHA system, introduced in 2023, has an operator move a pair of leader arms by hand while a pair of follower arms mirrors them, a setup known as leader-follower teleoperation. Its companion policy, Action Chunking with Transformers, learned fine-grained two-arm tasks from about 10 minutes of demonstrations per task.

Why it matters

Humanoid robots are built to work in spaces and with tools designed for two-handed people, so bimanual skill is central to whether they are useful. It is also harder to get right in the data. Both arms have to be recorded in sync, camera views multiply, with a wrist camera on each arm plus a head or overhead view, and different robots represent the same two-handed task with very different state and action vectors. That variation is a cross-embodiment problem, and it shows up the moment you compare datasets, as in 5 Humanoid and Bimanual LeRobot Datasets, Compared by State and Action Space.
Two arms double the joints and add timing a policy has to learn.

Frequently asked questions

What is a bimanual robot?

A robot with two arms, or two arms and hands, that it can coordinate on one task. Dual-arm industrial robots, research rigs such as ALOHA, and most humanoid robots are bimanual.

Why is bimanual manipulation hard?

The robot has to control twice as many joints, coordinate the timing between arms, and avoid colliding with itself, often while both hands touch the same object. The data is harder too, because both arms and several cameras must be recorded in sync.

What is ALOHA?

ALOHA is a low-cost, open source bimanual teleoperation system introduced in 2023, where an operator moves a pair of leader arms and a pair of follower arms copies them. It made collecting two-arm demonstrations affordable for many labs and led to follow-ups such as Mobile ALOHA.
black and white photo of Jesse Mostipak
Director of Growth
Jesse Mostipak is the Director of Growth at Voxel51, where the work is helping humans find and trust what the brand knows, and teaching the Google knowledge graph and the LLMs answering on their behalf to do the same. That question, how knowledge gets built inside a system, is one Jesse has been chasing for years. Earlier versions of it ran through a New York City high school science classroom, data science and machine learning, and developer relations at Kaggle, Posit (formerly RStudio), and Baseten. The answer doesn't change much depending on whether the learner is a teenager, a software engineer, or a knowledge graph. Jesse holds a Master's in Education from CUNY Hunter College. LinkedIn
See all articles by Jesse Mostipak
Last updated November 19, 2026

Building visual or physical AI?

Let's talk.