Inside the lab · 3 arms · 1 control interface

Same brain, different bodies.

One agent, three robot arms, one dashboard.

One hub, three arms running at once.

One column per arm: live cameras, progress, operator controls.

Task launcher: Piper, YAM and xArm running at the same time — Task launcher.

Real-robot eval dashboard.

Every trial, with synced video replay, the operator's verdict and a video-model judge side by side.

Three embodiments, same task.

Put the lego brick in the black box, run by GPT-6 Astra on each arm.

AgileX Piper · AgileX gripper

AgileX Piper puts the lego brick into the box

I²RT YAM · 4310 linear gripper

I2RT YAM puts the lego brick into the box

UFactory xArm · UFactory gripper

UFactory xArm puts the lego brick into the box

How it works.

Architecture: an operator station hub drives three Linux boxes, one per arm, over SSH tunnels; every arm shares one move_to control surface

More to come.

Next for the XEbench harness:

Dexterous handsMulti-finger hand tasks.
Two handsBimanual tasks.
More modelsVLAs (GR00T, π0) and world action models.
Human baselineTeleoperated runs on the same robots.
Sharper auto-judgeFailure modes and event timing from video.

Get the full technical report

Preliminary pilot data, not official Robotics Center or XEbench results. Robots, footage and data © Robotics Center. Model and robot names are trademarks of their owners.