WidowX 250 teleoperation: BridgeData-compatible data on a discontinued arm, and what replaces it

The WidowX 250 S is a 6-joint Interbotix arm built from DYNAMIXEL servos: 650 mm of reach, a 250 g working payload, 1 mm repeatability. It is also the robot behind BridgeData V2, 60,096 trajectories across 24 environments that a whole generation of generalist policies has been trained and evaluated on. Trossen has since discontinued it. That leaves two kinds of reader: labs that still run 250s and want more data that matches Bridge, and teams deciding what to buy instead.

WidowX 250, 250 S, 6DOF: which arm you have

The naming trips people up. Trossen's spec pages list the original WidowX-250 as a 5-DOF arm with eight servos, and the WidowX-250 6DOF (sold as the 250 S) as 6 DOF with nine, adding a forearm roll. Reach, span, payload and repeatability are the same. BridgeData V2 was collected on the 6DOF arm. If you are adding data to a Bridge-style project, check the joint count first, because a 5-DOF arm cannot reproduce 6-DOF wrist motion.

Two spec details matter for teleoperation. The 250 g payload is rated inside max reach, and Trossen recommends no more than 50% extension at that load, so long reaches with a heavy object strain the shoulder servos. And accuracy is listed at 5 to 8 mm against a 1 mm repeatability: the arm goes back to the same place reliably, but not necessarily to the place its model predicts. Teleoperation hides this, because the operator closes the loop by eye; scripted waypoints expose it.

How the WidowX 250 is teleoperated

Control rate: Bridge runs at 5 Hz, in end-effector space

This is the detail that matters most for compatibility. BridgeData V2 is recorded at 5 Hz, with an average trajectory of 38 steps. The LeRobot port, IPEC-COMMUNITY/bridge_orig_lerobot, makes the action space explicit: a 7-value end-effector action (x, y, z, roll, pitch, yaw, gripper) at 5 fps, with four 256×256 camera streams, across 53,192 episodes. That is very different from a modern leader/follower dataset recorded as joint positions at 30 Hz.

For remote operators, 5 Hz is forgiving: each step is 200 ms, so a 150 ms round trip is less than one step of delay, and Bridge-style pick-and-place in toy kitchens stays comfortable well beyond the latency a 30 Hz bimanual task can absorb. The VR stack, though, expects the headset on USB at the robot, so a remote setup has to carry the controller pose over the network to the robot host. Our latency guide covers the trade-offs.

Policy evaluation is a teleoperation job too

Because so much published work was trained or tested on Bridge data, a WidowX 250 is also a natural evaluation rig: a lab trains a policy and runs it on the same kind of arm, in a Bridge-like scene, to see how it compares. The rig matters less than the procedure around it. bridge_data_robot supports this with its server-client mode, where `widowx_env_service --server` runs on the robot host and the policy client runs on the same machine or another one. Evaluation still needs people. Someone sets each scene to the protocol, starts the rollout, stops it when the arm heads for trouble, and scores the result against a written success definition. That work suits a remote operator with a person on site for resets, and it benefits from the same discipline as data collection: identical scenes, identical scoring, and a log of every intervention.

Data formats and LeRobot

LeRobot has no built-in WidowX 250 driver: its built-in robots are the SO-series, Koch, OpenArm, OMX and a few others, and Trossen's LeRobot plugin registers WidowX AI types. Teams that want LeRobot recordings from a 250 write a small Robot class around the Interbotix SDK (connect, read joints, send targets), after which the usual LeRobot recording and training tools work. Bridge itself ships as raw JPEG, PNG and pickle files, with a preprocessed RLDS version for TFDS loaders. Other labs have published Bridge-compatible WidowX-250 data the same way, such as bridge_data_msr from Microsoft Research, 822 trajectories with three RGB views and depth.

We deliver WidowX episodes as LeRobot, RLDS or HDF5, and for Bridge-style projects we match the action convention and camera layout you specify. The data collection guide has the general QA checklist.

A Bridge-style collection protocol

BridgeData's value comes as much from how it was collected as from its size: many scenes, many skills, a language instruction per trajectory, and deliberate camera variation. If new data is meant to sit next to it, the protocol has to copy those choices, not only the robot.

What a good WidowX 250 operator does

Keeping a 250 fleet healthy

WidowX 250 discontinued: what replaces it?

Trossen's WidowX 250 page names no direct successor. Trossen sells two current lines: the WidowX AI and the newer WidowX Core, Pro and Heavy, which Trossen lists at 2, 4 and 6 kg of payload on Ethernet-connected arms with an iNerve controller, plus the Glide passive leader. Against the 250 S, that is a move from a USB servo bus and a 250 g payload to a different class of arm, recording joint-space leader/follower data at higher rates. Neither line is the same embodiment as BridgeData. If your evaluation depends on staying in Bridge's distribution, keep a 250 running for that work and collect new tasks on the new arms. For bimanual work, see our page on the Stationary AI, Trossen's current ALOHA-style cell.

How RoboPilots runs WidowX 250 data collection

Your arms stay in your lab. We agree the input (VR in the Bridge style, or a leader arm), set up the robot host on your side of a VPN, and train a dedicated group of operators on your unit and your task protocol, including the action convention, before the first recorded episode.