Buyer's guide: bimanual teleoperation kits for imitation learning (2026)
What a two-arm data-collection station consists of, what it costs to build versus buy, and what to check before a lab in the US orders one. Updated 2026-10-27.
What a bimanual data-collection platform is
A bimanual (two-arm) data-collection platform is a leader–follower teleoperation station in the layout made popular by ALOHA: two small leader arms that a person moves by hand, two follower arms that copy the motion, and cameras that record the scene and each gripper. Every episode produces synchronised joint positions, gripper commands and video, which is the training data for imitation-learning policies (ACT, Diffusion Policy) and for fine-tuning vision-language-action models.
Build or buy?
Labs that build their own typically combine four arms, a frame, two or three cameras, CAN adapters and a control PC, then spend weeks on wiring, calibration and timing. Buying an integrated station trades some of that flexibility for a known-good configuration: the arms are matched, the leader kinematics mirror the followers, the camera mast is at a repeatable height, and the SDK already supports both arms from one process.
| Item | Self-built (typical) | ARX Bimanual Platform |
|---|---|---|
| Follower arms | Chosen separately; payload and force sensing vary | 2 × ARX X5: 6-DoF + gripper, torque sensing at every joint, about 1 kg peak payload, about 780 mm span |
| Leader arms | Often 3D-printed or a second pair of arms | 2 matched leader arms |
| Cameras | Buy and mount yourself | Top mast camera + wrist cameras on the base |
| Software | Assemble drivers per arm | Open-source arx5-sdk: both arms from one PC, 500 Hz, joint and Cartesian control |
| Price | Varies widely with the arms chosen | $5,999 launch price for the first 20 units, then $6,999; plus freight, shipped from the US |
What to check before ordering
- Force sensing. Torque sensing at every joint lets you record contact-rich tasks and use compliant control when deploying policies; position-only arms cannot.
- Payload at the gripper. About 1 kg peak per X5 arm covers tableware, tools, cloth, food packaging and most lab manipulation benchmarks. If you need several kilograms, look at a larger arm instead.
- Control rate and latency. 500 Hz joint control with sub-millisecond motor communication is what you want for smooth teleoperation and for running learned policies at 50 Hz or more.
- Open SDK. Check that the arms are supported by code you can read and modify. arx5-sdk is on GitHub, installs with pip and already has a two-arm example.
- Models. URDF and STEP files make simulation, collision checking and custom mounts straightforward.
- Where it ships from. Robotics hardware imported into the US usually carries duties, broker fees and multi-week freight; a unit stocked in the US arrives in days with none of that.
- Procurement path. Universities generally need a formal quote, a W-9 and an invoice against a purchase order. Confirm the seller can do this before the fiscal-year deadline.
How the ARX platform fits
The ARX Bimanual Platform is the integrated version of this layout: two X5 followers, two leaders, mast and wrist cameras on one base, driven by open-source Python. ARX Global stocks it in the United States and sells it at a $5,999 launch price for the first 20 units (then $6,999) plus freight, with card checkout or purchase-order invoicing.
See the platform Request a quote
Sources
- ARX store listing (bimanual data-collection platform): arx-x.com
- arx5-sdk README (Stanford REAL Lab): github.com/real-stanford/arx5-sdk
- ARX X5 SDK and models: github.com/ARXroboticsX
- X5 figures (payload, span, weight) as confirmed by ARX, October 2026; see arx-global.us/hardware/x5