In-house research · Simulation only
G1 box transport in a custom factory simulation
Integrating a walking policy and two-handed support control to move a box from a shelf to a tray. Three successes in three fixed-layout trials; one successful placement in ten shifted-position trials.
What we built
We built a custom factory simulation in which a Unitree G1 humanoid removes a box from a shelf, walks along an aisle, places it on a tray and releases it. TeamZ integrated the environment, route, hand support positions and controllers. We did not collect additional teleoperation demonstrations for this work.
The walking policy is combined with two-handed support control. Near the tray, the robot switches from walking to standing control before lowering the box. We adjusted hand insertion depth and this transition after early trials showed box tilting and backward movement during placement.
Results and acceptance conditions
The fixed layout succeeded in all three trials, restarting the simulator each time. In ten conditions with the box's initial position shifted by up to ±2 cm in each horizontal direction, placement and settling succeeded in one trial. One trial also met the additional contact and tilt conditions. These are results for this setup, not a success-rate guarantee for other boxes, shelves or routes.
Of the nine unsuccessful trials, seven exceeded the 20 cm movement limit during placement, one exceeded the 30 cm arm-position correction limit and one exceeded the time limit. Failed and interrupted runs are included in the results. The ten shifted positions used random seed 42.
Placement required the entire box to remain inside the tray, near its support surface, at less than 0.05 m/s for at least three seconds. Additional criteria were two-hand contact in at least 99% of transport control frames, a maximum box tilt of 30 degrees, and no hand contact during the final three seconds.
In this transport experiment, open hands support the box rather than closing the fingers around it. Recorded contact forces do not establish guaranteed holding strength. Trials are stopped for unsafe states within the simulation, including near-falls, drops and excessive movement during placement.
How the motion is controlled
End-effector targets from public demonstration data are adapted to the shelf height and layout. Controllers use those targets while the simulator calculates gravity and contact. The box is neither teleported nor attached to the hands during transport.
The source dataset contains human demonstrations. Avoiding new teleoperation collection at TeamZ does not mean that no human demonstration data was used. Locomotion uses a pretrained policy; removal, transport and placement use scripted control. ACT training had not been performed for this custom transport task in this experiment.
Development update: grasping and holding
In a separate basic test on September 8, 2026, we revisited grasping and holding with a different controller configuration, using Isaac Lab-Arena and AGILE/PINK. Closing the fingers lifted the bottom of the box at least approximately 8.8 cm above the shelf, with both hands maintaining contact for the final three seconds.
The test did not pass the stable-hold criterion: maximum box speed was approximately 0.23 m/s, above the 0.08 m/s limit. This was a scripted simulation test before ACT training, not a completed transport result. Contact and wrist motion were recorded to investigate the oscillation.
What this demonstrates—and its limits
This work connects simulation environment construction, robot control and repeatable evaluation. We record hand contact, box pose, tray containment and settling after release rather than judging success from the video alone.
The transport shown here has not been validated on a physical robot. Simulation helps identify conditions to check before hardware deployment; it does not establish real-world reliability. TeamZ can work with your development team on simulation environments, robot control, learning-data preparation and evaluation.
Environment and attribution
Transport experiment: Unitree G1 with Dex3, Isaac Sim 6.0.0-rc.22, Isaac Lab-Arena, Unitree's public walking policy, and PINK/HOMIE. The original handling motion uses NVIDIA's Arena-G1-Loco-Manipulation-Task public dataset under CC BY 4.0. TeamZ adjusted the factory environment, route, support positions and controller connections.
Sources
LET’S BUILD TOGETHER
What are you working on?
Tell us about the software you need to implement or test. We can discuss the work, the development setup and how we could contribute.
Discuss your development needs