Single-Arm vs Bimanual Manipulation Data

Try folding a bath towel with one hand. You cannot. One hand pins a corner while the other sweeps the fold flat, and the moment you remove either, the task collapses. A large share of everyday manipulation is like this: not two hands doing twice the work, but two hands doing one job that neither could finish alone.

Most robot manipulation datasets are single-arm. That is a quiet, consequential omission. It does not just leave out a few tasks; it leaves out an entire category of human competence, and it does so precisely where the data is hardest to collect.

The instinct is to treat bimanual capture as single-arm capture, doubled. It is worse than that, and understanding why is the difference between a dataset that teaches coordination and one that only looks like it does.

Why bimanual is more than twice the problem

Add a second arm and the action space doubles. A seven-joint arm becomes fourteen joints, plus two grippers. But dimensions are the easy part. The hard part is coupling: the two arms are not independent, and a policy has to model how they constrain each other in time and in contact.

Consider what single-arm data never contains. Two end-effectors that must arrive at the same place at the same instant. An object held by one hand while the other works on it. Forces passed between the arms through a shared rigid or deformable thing. None of that appears when you log one arm at a time, so none of it can be learned from single-arm logs. Work on bimanual policies from Physical Intelligence keeps circling this coordination structure as the core difficulty.

The math compounds in a way that is easy to miss. If a task needs the two hands in a particular relative configuration, the space of joint configurations the policy must cover is closer to a product than a sum. Two arms do not span twice the states of one. They span something much larger, and coverage is what demonstrations buy.

The coordination tax

Two-handed tasks are not all alike. They fall into a few patterns, and each asks something different of the data.

• Symmetric. Both hands do mirrored work, like lifting a large box. The demand is tight spatial and temporal symmetry.
• Asymmetric. One hand stabilizes while the other acts, like holding a jar and twisting the lid. The stabilizing hand is easy to under-record and easy to underrate.
• Sequential handoff. An object passes from one hand to the other. The transfer instant is a moment of shared contact that must be captured on both sides at once.

Every one of these breaks if the two arm streams drift apart in time. A misalignment of a few tens of milliseconds between the arms can turn a clean handoff into a drop. This is why temporal synchronization stops being a nicety and becomes the thing the whole recording depends on.

Notice how different these patterns are as data. A symmetric lift can tolerate loose timing. A handoff cannot; miss the shared-contact window and the object is on the floor. A dataset that mixes them without labeling which is which hands the policy an averaged, blurry target.

Single-arm errors stay local; in bimanual work, one hand's mistake becomes the other hand's problem a fraction of a second later.

What bimanual capture actually demands

The requirements compound. You need two arm streams synchronized to the millisecond, not merely two recordings that start together. You want a wrist camera on each side, because each hand can occlude the other from any fixed view. You need to capture the contact between the arms, not just each arm's contact with the world. And because the coordinated state space is so much larger, you need many more demonstrations to cover it than a single-arm task of the same nominal difficulty.

The public corpora reflect how hard this is. Large aggregations like Open X-Embodiment, single-arm collections like the DROID dataset, and contact-rich sets like RH20T are dominated by single-arm data, because single-arm data is far cheaper to collect at scale. Bimanual demonstrations remain comparatively scarce, which makes them disproportionately valuable.

There is a teleoperation cost too. Collecting bimanual data by teleoperation means one operator driving two arms at once, which is harder, slower, and more tiring than driving one. Throughput per operator drops, so the scarce data is also the expensive data. That is a double squeeze on supply.

The data gap, and who is placed to close it

Here is the twist. The scarcest data in robotics is the most natural data for a human to produce. People are bimanual by default. We stabilize and act, hand off, and coordinate two hands thousands of times a day without thinking about it. A single-arm teleoperation rig cannot express any of that; a person simply does it.

That makes first-person human demonstration structurally well-suited to the bimanual gap, provided the capture actually preserves the coupling: both hands, synchronized, with contact and viewpoint recorded on each side.

This is the quiet reason humanoids matter for data, not only for deployment. A two-armed, two-handed platform, or a person wearing capture gear, can produce coordinated demonstrations that a single-arm bench simply cannot represent. The embodiment shapes the data it is able to record.

So single-arm versus bimanual is not a slider from easy to hard. It is a change of kind. The action space doubles, but the coordination, contact, and timing requirements more than double on top of it. Datasets that ignore this will keep producing robots that pick and place beautifully and cannot open a jar. The ones that capture two hands as a single coupled event are building toward the tasks that actually fill a human day.