What Is ALOHA (Bimanual Teleoperation)?

ALOHA is introduced in the paper Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware (arXiv 2304.13705) by Zhao, Kumar, Levine and Finn, with an official project page and the ACT GitHub repository. The authors ask whether learning can let low-cost, imprecise hardware perform fine manipulation tasks such as threading cable ties or slotting a battery, which they say usually needs high-end robots, accurate sensors or careful calibration.

ALOHA is third-party research, and Firsthand is not part of it, endorsed by it or affiliated with its authors. Firsthand makes no claim that its data trained or was used to evaluate ALOHA or ACT, and no performance claim against them. The site does not state that Firsthand offers teleoperation, and Firsthand does not capture robot action data. What Firsthand states is custom egocentric human capture: both hands in frame with per-hand 3D pose and visibility flags, plus depth, contact events and action segments, collected fast, on demand and to a written skill spec.

The project page defines ALOHA as a low-cost open-source hardware system for bimanual teleoperation, and says a $20k budget is enough for tasks such as threading a zip tie and juggling a ping pong ball. The paper says demonstrations are collected with a custom teleoperation interface.

Policy observations come from four RGB cameras, each streaming at 480x640: two stationary and two mounted on the robots’ wrists. The interface’s arm layout, degrees of freedom and bill of materials are Not stated in the sources read.

ACT predicts a sequence of actions, an action chunk, instead of a single action as in standard behavior cloning, which the project page says reduces the effective horizon. The policy is trained as the decoder of a conditional VAE: a transformer encoder takes images from multiple viewpoints, joint positions and a style variable, and a transformer decoder predicts the action sequence. At test time the CVAE encoder is discarded and the style variable is set to zero.

On the project page the policy directly predicts joint positions at 50Hz with a fixed chunk size of 90, against episode lengths of 600 to 1000.

A demonstration pairs the camera streams with joint positions, which ACT takes as input and predicts as output. The ACT repository reads episodes from an .hdf5 dataset and includes two simulated environments, Transfer Cube and Bimanual Insertion, with scripted policies; its example generates 50 episodes. Real-world field lists, episode counts and RLDS or LeRobot export are Not stated in the sources read.

Because the actions are the robot’s own joint positions, the data matches that robot exactly, a trade the teleoperation entry calls slow and hardware-bound.

The abstract reports that ACT learned six difficult real-world tasks, such as opening a translucent condiment cup and slotting a battery, with 80-90% success from only 10 minutes of demonstrations. The project page shows rollouts from 50 demonstrations per task, with 96%, 84%, 64% and 92% success on four tasks.

The README says simulated success should be around 90% on transfer cube and 50% on insertion. These are the authors’ results on their own setup, and Firsthand draws no comparison.

ALOHA data comes from a person driving a robot, so the robot is needed during collection. UMI replaces the robot with a handheld gripper, and its entry reports the paper finding the gripper more than 3 times faster than space-mouse teleoperation on one task. Egocentric human capture uses the contributor’s own hands, which leaves an embodiment gap to close, per the teleoperation entry. That entry says most programs need a blend.

Firsthand ranks none of them.

Write the target bimanual skills, environments and signals into a skill spec and confirm them on a Spec Pilot, a small fast first batch. This is on-demand custom collection of egocentric human data, not a teleoperation program and not a catalog download.

Where coverage is deep, the median from signed spec to first validated batch is 48 hours; where it is thin, three to four weeks. Billing is per validated hour.

What is ALOHA?

A Low-cost Open-source Hardware System for Bimanual Teleoperation, from a 2023 paper.

What is ACT?

Action Chunking with Transformers: a policy that predicts a chunk of actions instead of a single action.

What data does ALOHA produce?

Camera streams paired with joint positions, in .hdf5 episodes. Real-world field lists are Not stated in the sources read.

How is ALOHA different from egocentric human data?

ALOHA data comes from a person driving a robot and matches that robot exactly. Egocentric human data leaves an embodiment gap to close.

Does Firsthand offer teleoperation?

The site does not state that it does. Firsthand states custom egocentric human capture, delivered in RLDS, LeRobot, WebDataset or HDF5 / zarr.

Is Firsthand data part of ALOHA?

No. Firsthand is not part of it or affiliated with its authors, and makes no claim that its data trained or evaluated ALOHA or ACT.