The suite is headlined by the Trossen Workbench and Rivet, its stationary and mobile bimanual manipulation platforms. Physical AI models are only as good as the demonstration data they are trained on and most of that data is visual. Today, teams typically assemble their own rigs from individually sourced robot arms and consumer-grade USB cameras, relying on custom driver code to connect them. The result is datasets with low-resolution imagery, inconsistent calibration, motion blur, and timing drift across views and data streams.

The Trossen Workbench and Rivet are unified development platforms featuring dual WidowX Pro 6-DoF arms with reach from 700 mm to 1000 mm, payload capacities of 4 to 6 kg payload, and 1 mm repeatability. Each platform is powered by an onboard NVIDIA Jetson AGX Orin 64 GB compute and supports teleoperation over local networks and the internet. Each also includes a factory-calibrated three-camera vision system built entirely with Stereolabs hardware.

The Stereolabs and Trossen joint vision architecture mirrors how state-of-the-art manipulation policies are trained. A center-mounted Stereolabs ZED X Mini provides the global scene view and stereo-depth context of the full workspace, while each wrist-mounted ZED X Nano delivers the close-range view of the gripper and object that imitation-learning policies rely on for fine manipulation.

The ZED X Nano was designed specifically for this application. Its dual 2.3 MP (1920×1200) global-shutter sensors capture at up to 60 fps without the motion blur of rolling-shutter USB cameras, and its neural depth engine resolves geometry from as close as 3 cm, the distances at which grasping actually happens. GMSL2 connectivity with locking, EMI-resistant cabling keeps all three cameras deterministically synchronized on the Jetson with a zero-copy pipeline, so teams can record, encode, and run inference simultaneously without frames silently dropping mid-episode. A vibration-resistant onboard IMU keeps the wrist views usable even as the arms move at speed.

Every teleoperated episode collected on the Workbench and Rivet is a clean, synchronized, multi-view, RGB-plus-depth training sample, ready for imitation learning, reinforcement learning, and sim-to-real workflows through native support for ROS 2 and NVIDIA Isaac Sim / Isaac Lab.