ALOHA is less a single product than a lineage: a bimanual, leader-follower teleoperation rig that came out of Stanford in 2023, the imitation-learning policy (ACT) that made its data unusually sample-efficient, and a growing set of derivatives — Mobile ALOHA, ALOHA 2, and commercial kits — that all trace back to the same core idea. If your data collection program touches bimanual manipulation, you have almost certainly seen an ALOHA-shaped rig in a paper or a product page, whether or not you use the name.
The core insight was hardware-economic rather than algorithmic: fine, contact-rich manipulation had mostly been demonstrated on arms costing well into six figures. ALOHA showed that a puppeteered, backdrivable leader-follower pair costing on the order of $20,000 — orders of magnitude cheaper — could collect demonstrations good enough to train policies that threaded cable ties and slotted batteries.
Control interfaces that matter for data collection
The reference ALOHA rig pairs two Trossen Robotics WidowX arms as leaders with two ViperX arms as followers, each 6 DoF, with the WidowX sharing the same kinematic structure as the ViperX in a smaller, lighter form factor built for backdriving. Follower joint targets are simply the leader's measured joint positions — there is no inverse kinematics or retargeting layer in the reference design, which keeps the control loop simple and low-latency.
That simplicity is deliberate. Because both arms share a kinematic structure, "backdriving" the leader produces a follower trajectory with no interpretation error: what the operator's hand does, the follower's joints do, sample for sample. This is the textbook case of bilateral teleoperation — the operator's motor loop and the robot's joint-space trajectory are the same loop.
Mobile ALOHA (Fu, Zhao, Finn — Stanford, posted to arXiv in January 2024 and presented at CoRL 2024) keeps this arm architecture and adds an omnidirectional mobile base plus a whole-body teleoperation interface, so the same operator controls base motion and both arms together rather than driving the base and the arms as separate systems. The paper's headline result is about data efficiency across embodiments as much as mobility: co-training on the existing static ALOHA datasets alongside a small number of mobile demonstrations was reported to increase task success rates substantially — one more reason to treat ALOHA-family data as a connected corpus rather than isolated per-rig datasets when you can.
ALOHA 2, the Google DeepMind-led hardware revision published in 2024 with Stanford and Hoku Labs collaborators, kept this architecture but reworked the gripper mechanism and mounting hardware — reported improvements include roughly a 10x reduction in the force an operator needs to apply at the gripper and roughly double the follower's gripping force, aimed specifically at reducing operator fatigue across long fleet-scale data collection sessions.
Teleoperation options
Puppeteering the leader arms is the only interface the reference hardware supports, and that is the point — it is what makes ALOHA data look the way it does. Two consequences follow directly:
- No retargeting error. Because leader and follower share kinematics, every demonstration is a clean joint-space trajectory. Compare this to VR-controller or hand-tracking interfaces, where a retargeting layer between operator pose and end-effector pose introduces its own noise (see VR vs. leader-follower teleoperation).
- No force feedback beyond mechanical backdrive. The operator feels the follower's real inertia and friction only through the leader's own linkage, not through active haptic rendering. It is closer to moving a lightweight physical proxy than to a force-reflecting bilateral system.
Bimanual coordination is where ALOHA earns its name over single-arm rigs: tasks like transferring an object hand-to-hand, or stabilizing with one arm while the other manipulates, are captured naturally because both arms are puppeteered simultaneously by the same operator, two-handed.
What a clean dataset looks like on this platform
ALOHA-style datasets record, per timestep and per arm: leader joint positions (the action), follower joint positions (the observation.state), gripper state, and multiple camera streams — the original design specifies overhead, wrist-mounted, and worm's-eye viewpoints so the policy has enough visual context to resolve the fine manipulation the platform is built for.
- Follower arms
- 2x ViperX, 6 DoF each
- Leader arms
- 2x WidowX, 6 DoF each, backdrivable
- Cameras (reference build)
- Multiple viewpoints incl. wrist-mounted
- Companion policy
- ACT — action-chunked transformer
ACT itself shapes what "clean" means for this data: because the policy predicts action chunks rather than single steps, the training signal is more forgiving of small per-step teleoperation noise but more sensitive to consistent chunk-length dynamics — an operator who pauses mid-motion, then resumes, creates a discontinuity that a chunk-based policy has to learn as a distinct mode rather than smooth it away.
Common pitfalls
Undertraining the operator on bimanual coordination. Two-handed puppeteering has a real learning curve distinct from single-arm teleoperation — coordination between the arms, not just precision within each, is part of what the recording is supposed to capture. Rushed operators produce demonstrations where the "off-hand" arm is doing something incidental rather than task-relevant, and that noise trains directly into the policy. See teleoperator training.
Assuming Mobile ALOHA's whole-body interface ships with the stationary rig. Mobile ALOHA adds a mobile base and a distinct whole-body teleoperation scheme layered on top of the original bimanual arms — it is a different control problem (base velocity plus two arms, reported as a combined action space) from stationary ALOHA, not a firmware update to it.
Pooling data across ALOHA hardware revisions without noting it. ALOHA 1 and ALOHA 2 differ in gripper mechanics and operator force requirements, which measurably changes demonstrator behavior even when the kinematic structure is unchanged. Tag which hardware generation each episode came from.
Treating ACT as required. ACT is the policy ALOHA was validated with, not a constraint on the hardware. ALOHA-format datasets — bimanual joint trajectories plus multi-camera video — train other imitation-learning architectures fine; nothing about the MCAP-or-HDF5-adjacent recording layer is ACT-specific.
Underestimating operator fatigue on the original ALOHA gripper. Part of what ALOHA 2's redesign targeted was exactly this: the original gripper mechanism required enough hand force from the operator that long fleet-scale sessions were tiring in a way that shows up as degraded demonstration quality late in a session. If you are running an ALOHA-1-generation rig for extended collection, budget operator breaks accordingly, or account for the ALOHA 2 hardware revision if throughput matters.
Whether you build a reference ALOHA rig, buy a commercial bimanual kit, or run a different bimanual platform entirely, the recording discipline is portable: bring your own rig covers wiring a bimanual arm pair into the same time-synchronized capture pipeline used for any other hardware.