Trossen WidowX AI teleoperation: leader arms, the LeRobot plugin and data OpenPi can train on

The WidowX AI is Trossen Robotics' arm for machine-learning research: 6 joints plus a gripper, a 1.5 kg payload, 1 mm repeatability and a 4 kg body, sold in Base, Leader and Follower configurations. It is the arm behind Trossen's single-arm Solo AI kit, and it comes with a maintained LeRobot plugin and an OpenPi integration. That makes it one of the shortest paths from an empty bench to a trained ACT, smolVLA or π0 policy. The hard part is not getting it to move. It is recording episodes that are consistent enough to train on, at a rate a team can sustain.

Solo AI, a standalone pair, or a bimanual kit

Trossen describes the Solo AI as "a compact, single-arm research platform for machine learning in the field" that "can be used for teleoperation, recording datasets, and testing inference". It is listed at $11,385.95 on Trossen Robotics as of October 2026. Trossen says Solo AI "is designed for data collection in the field and for those NOT needing bimanual manipulation". For two-arm work it sells bimanual kits built on the same arm: the Stationary AI for a controlled lab, where arm and camera placement stay the same between sessions, and the Mobile AI on a wheeled base.

If you already have cameras, compute and a table, buying a single leader plus a single follower gets you the same arm hardware. The Follower configuration includes a RealSense D405 on an arm mount; the Leader swaps the gripper for an ambidextrous hand grip with sliding-rail finger pinchers, which is what the operator holds.

Three ways to drive a WidowX AI

The driver also ships a `teleoperation` demo that Trossen describes as teleoperation "with force feedback", alongside Cartesian impedance and admittance examples. LeRobot's leader/follower loop does not use force feedback, so treat the demo as a starting point if contact feel matters for your task, not as part of the standard recording path.

The control loop: 30 Hz by default, and what that means remotely

Under the hood, the arm controller takes real-time commands and returns state over UDP, with TCP for the handshake, configuration and logs. Each joint can run in idle, position, velocity, external-effort or raw effort mode. Commands carry a goal time: above 0.2 s the driver uses quintic interpolation for positions, between 0.001 s and 0.2 s it interpolates linearly, and at 0.001 s or less the goal applies immediately. Leader/follower streaming lives in that short-goal-time regime.

In the LeRobot plugin the follower's `loop_rate` defaults to 30 Hz, and Trossen's camera examples run at 640×480, 30 fps. Two other defaults shape how teleoperation feels: `max_relative_target` (5.0 on the single-arm follower, unset on the bimanual one) caps how far a single step can move, and `min_time_to_move_multiplier` (3.0) sets the minimum move time as the multiplier divided by fps, which is 0.1 s at 30 fps. Raise the multiplier and motion gets smoother but lags; lower it and motion gets faster but can turn jerky. If your policy will run at 20 Hz, decide up front whether you record at 20 Hz or record at 30 Hz and resample, and keep it the same for the whole dataset.

At 30 Hz a control step is about 33 ms. An operator on a 100 ms round trip is three steps behind the camera; at 200 ms, six. Tabletop transfer, sorting and handovers stay workable well into that range. Tight insertions and cloth manipulation do not. Two practical rules for remote operation on this arm:

Our latency guide has budgets by task type. Note too that OpenPi inference is a different loop: Trossen's tutorial runs control at 30 Hz to match the cameras and warns that querying the policy too often can drop the effective control rate to around 10 Hz.

Data: LeRobot v3 for recording, LeRobot v2 for OpenPi

The plugin requires `lerobot >= 0.6.0`, auto-registers its robot and teleoperator types, and writes standard LeRobot datasets. Trossen's single-arm example records 10 episodes of 45 s with a 15 s reset; the bimanual example records 25 episodes of 60 s. Velocity, effort and external-effort observations are off by default (`include_velocity`, `include_effort`, `include_external_effort`). Turn on the ones you want before the first episode, because adding a field later splits the dataset in two.

For a concrete reference, TrossenRoboticsCommunity/widowx_ai_orange_block_transfer is a bimanual set in LeRobot format v3.0: 20 episodes, 13,180 frames at 30 fps, a 14-value state and action (six joints plus the gripper carriage per arm), and four AV1 camera streams (high, low and one per wrist). The same Hugging Face organization publishes ACT, smolVLA, π0 and π0.5 checkpoints trained on WidowX AI data, such as smolvla-handover-cube-widowxai and pi0-handover-cube-widowxai.

We deliver episodes in LeRobot, RLDS or HDF5, and the same rules apply whichever you pick: one action convention (joint targets) for the whole set, one clock, a fixed start pose, and an honest success flag. The data collection guide goes deeper on checks.

What a good WidowX AI operator does

That last point is about staffing as much as skill. Trossen's example reset window is 15 s. A remote operator can only hold that pace if someone near the bench puts objects back, swaps parts and clears drops. Depending on the task, one technician may be able to cover more than one station; plan for that role from day one.

WidowX AI, WidowX Core/Pro/Heavy or xArm 6?

Against the newer Trossen line, the WidowX AI has the deeper software: a released LeRobot plugin with single, bimanual and mobile types, OpenPi recipes, and public datasets and models. The Core, Pro and Heavy add payload (2 kg, 4 kg and 6 kg), a controller Trossen says runs above 600 Hz, and the Glide passive leader. Against a UFACTORY xArm 6, the WidowX AI has a far smaller payload and a looser 1 mm repeatability, but it is lighter, ships as a matched leader/follower pair and records through a maintained plugin, where the xArm relies on community or UFACTORY integrations.

How RoboPilots runs WidowX AI teleoperation

The arms stay in your lab. We agree the input with you (your WidowX AI leaders or Trossen VR), set up the recording host on your side of a VPN, and train a dedicated group of operators on your unit and your task protocol before the first recorded episode.