What happens when we reduce friction in G1 box transport?

We tested G1 lifting, carrying and placing a 1 kg box in a custom factory environment in Isaac Sim. The original friction condition completed the task, while four lower-friction conditions failed during lifting. We show both success and failure.

Following the flat-ground experiment, this test ports the setup to Isaac Sim and compares contact conditions. Each condition uses the same initial placement and one trial. This does not demonstrate physical-robot transport.

The condition that completed transport

We used the public pretrained HAIC policy and reference motion in Isaac Sim 5.1 and Isaac Lab 2.3. The original model’s whole-body collisions remained enabled; collisions were not limited to selected contact parts.

The original condition passed ten checks covering lifting, transport, placement and hand release. Maximum box-bottom height was about 65 cm and final position error about 9 mm.

Lower friction prevented the lift

We changed the box’s friction settings while retaining the policy, box mass, initial placement and foot/floor materials. The four conditions at effective friction 0.4 or below failed to lift sufficiently and ended on the robot-fall criterion.

At zero friction the box slid across the floor. Displacement alone is not counted as success: lifting, placement position, rest and release are checked separately.

Material settings are not the effective contact friction

The original box coefficient is 1.0 and the robot coefficient 0.5. PhysX’s average combine mode produces effective friction of 0.75. The comparison conditions change the box’s combine mode to minimum. A box setting alone does not describe friction at contact.

Box–floor friction changes at the same time. This does not isolate hand–box friction. One trial per condition cannot determine a failure boundary or a general success rate.

Scope and limitations

TeamZ did not collect additional teleoperation data or retrain the model. This does not mean the original training used no human-derived data. The policy uses reference motion and simulator state; it does not recognize the box through a camera and navigate to it.

The videos re-render poses saved during physics tests at 25 fps and normal simulation speed. Pose replay for presentation is separate from policy control during the tests.

The box is a 1 kg rigid body, and hand/forearm collision shapes are simplified. Articulated five-finger hands, a 2 kg box, real cardboard friction and deformation are untested. Friction values were not calibrated from physical measurements.

Next, we plan to vary hand friction while holding box–floor conditions fixed, then separate the effects of holding posture and speed. The aim is to establish operating conditions as well as visible motion. Discuss robotics simulation and implementation with TeamZ.

Primary sources

HAIC research page · HAIC paper · HOIC-baseline execution code · HAIC training code