Watch G1 sort four items in 73 seconds

An internal warehouse simulation. The left view follows the fingers. The right shows the full body. Each item traveled 1.72–1.80 m and stayed still in its assigned bin for 0.5 s after release.

Clearance for the book

In an early trial the robot grasped the book, then knocked the neighboring bottle off its stand as it stepped sideways. We changed the controller to wait until both the fingertips and the bottom of the book cleared the bottle. The film shows the box, block, book and bottle sorted in one continuous run.

What the robot learned

Unitree’s gait model controls the legs. A controller closes the fingers and checks contact. We trained a PPO policy to adjust travel velocity by up to 20% and left-arm joint targets by up to 0.01 rad during carry. An image classifier selects a destination from a simulator-masked crop. Shelf coordinates are supplied.

The revised reward gives 2 points for lift, 3 for carry, 1 for release and 10 for settling in the correct bin. Time and shoulder torque incur a cost. A separate model trained with this reward completed 5 of 8 single-item trials with shifted starts (4 cm per axis), versus 4 of 8 without the learned corrections. Both block trials failed with the learned policy.

Shoulder load

The SDK reports position, velocity and torque for 29 body joints. Left shoulder pitch torque peaked at 11.48 N·m while grasping the box. Predicting motor temperature requires calibration against current and temperature readings from hardware.

Synthetic data for robot AI development

We recorded 172 synchronized sets of RTX-rendered RGB images, item and shelf masks, camera data and point clouds. The image classifier used masked RGB crops, evaluated with episode-level splits. Separately, a PPO policy adjusts walking commands and left-arm targets; a controller checks finger contact during grasping. The point clouds use PhysX distance rays.

Picking a carton from a mobile shelf

A mobile base delivers the shelf. A seven-axis Franka Panda transfers a carton into a tote. We approximate suction by adding a constraint after cup contact and releasing it over the tote. In this simulation the carton rose 20.5 cm and settled 5.3 mm from the target.

Discuss a task to simulate

A shelf drawing, item dimensions and an example of a failed operation help define the first trial.

Contact us ↗