Mobile ALOHA teleoperation on the Trossen Mobile AI: bimanual data away from the lab bench

Mobile ALOHA took the ALOHA bimanual cell off the bench: the project "augments the ALOHA system with a mobile base, and a whole-body teleoperation interface", learned tasks such as cooking shrimp, storing pots in a wall cabinet and calling an elevator from 50 demonstrations each, and showed that co-training with static ALOHA data helped. Trossen's current version is the Mobile AI, "a high-performance, mobile all-in-one robotic machine learning workstation" built from two WidowX AI leader/follower pairs, three RealSense D405 cameras and a touchscreen on a wheeled base. Its pitch is moving between labs, warehouses and other sites without disassembly. That is also what makes it harder to operate than its stationary sibling.

What changed from the original Mobile ALOHA kit

Trossen's ALOHA 2.0 specifications describe the earlier Mobile ALOHA on a SLATE base: 122 kg, 1 m/s top speed, adjustable training wheels and a 1.4 kWh portable power station, with ViperX followers and WidowX leaders. In that kit's operation guide, one person teleoperates the arms "while doing teleop, another person can use the included PS4 controller to move the base", holding L2 to enable driving and releasing it to stop.

The Mobile AI swaps both arm types for WidowX AI pairs and records through Trossen's LeRobot plugin as `mobileai_robot` with `mobileai_leader_teleop`. Trossen's product page describes an integrated wheeled base and lists the SLATE base as compatible for autonomous navigation workflows. The plugin's Mobile AI config adds one option on top of the bimanual one, `enable_base_motor_torque`, which defaults to False. If you own a legacy kit, its ROS-based scripts and HDF5 archives still work on their own terms; the step that needs care is merging them with new data, covered below.

Driving it: the leader arms and the people around them

Arm teleoperation on the Mobile AI is the same joint-for-joint leader/follower setup as the Stationary AI. The teleoperation docs give each of the four arms its own IP address, three RealSense cameras (`cam_high`, `cam_left_wrist`, `cam_right_wrist`) at 640×480 and 30 fps, a `max_relative_target` safety limit of 5.0 that Trossen suggests removing only "once you're confident in controlling the robot safely", and a 30 Hz loop rate. Ctrl+C stops it.

What changes is the room. A stationary cell has a fixed table and fixed lighting. A mobile kit goes into a kitchen, a stockroom or a corridor, and every new site needs someone there to move the kit, keep people clear of the arms, and put the scene back between attempts. Plan the Mobile AI as a two-role job from the start: the operator on the leaders, and a field technician who handles the base, the site and the resets. Legacy Mobile ALOHA made the same split explicit with its second person on the PS4 controller. Over a long day the technician also watches cable strain on the wrist cameras, power, and anything on the floor the wheels could catch.

Control rate and latency in the field

The recording loop defaults to 30 fps, so each step is about 33 ms. In a lab, the network between the operator and the robot host is a known quantity. In the field it is whatever the site has: shared Wi-Fi, a hotspot, a warehouse network with client isolation. Three rules keep remote operation on a Mobile AI sane:

Trossen's recording docs also note that camera fps can become unstable, and suggest more image-writer threads per camera, turning off the live display, or switching the RealSense cameras to the OpenCV interface. On a laptop-powered kit, check the recorded fps of every episode, not just the first.

Data: what to record, and how much

The plugin writes LeRobot datasets. Trossen's Mobile AI recording example sets 60 s episodes with a 15 s reset (the plugin's own default reset is 60 s), with joints in radians and the gripper in meters. The public Trossen T-shirt folding set gives a sense of scale for a real task: 101 episodes and 155,841 frames at 30 fps, three AV1 camera streams, in LeRobot format v3.0. Trossen's product page lists the optional laptop as optimized for ACT and ACT++ training and evaluation on the kit; for π0 and π0.5, Trossen's OpenPi recipe has its own dataset-version requirements, covered on our WidowX AI page.

A field day, step by step

Most lost episodes on a mobile kit are lost to logistics, not to operator skill. A fixed routine for each site visit catches the usual problems before they reach the dataset:

What a good Mobile AI operator does

Mobile AI or Stationary AI?

Buy the Mobile AI if the task truly needs to happen in different places: homes, stores, warehouse aisles. Buy the Stationary AI if it can happen at a bench, because a fixed frame with four cameras gives you more repeatable episodes for less effort per episode. Both use the same WidowX AI pairs and the same plugin, so operator training and task protocols carry over. If you are weighing Trossen's newer arms instead, the WidowX Core, Pro and Heavy line, which pairs with the Glide passive leader.

How RoboPilots runs a Mobile AI program

The kit stays with you, at your sites. We agree the input path and the network setup, put the operator client on your side of a VPN, and train a dedicated group of operators on your unit, your sites and your task protocol before the first recorded episode. You provide the field technician who moves the kit and resets the scene; we write the reset checklist with you.