Testing the physics of G1’s two-handed valve manipulation

This is a preliminary simulation experiment, not verification that a physical G1 can open or close a real valve. We applied opposing torque while G1 turned a handle with both hands. Four Isaac Sim conditions from 0 to 0.6 Nm were compared. At 0.6 Nm, rotation reached 45.88 degrees; load remained applied after stopping while we checked both-hand contact and standing.

Can it hold against a load?

A known torque opposes opening. It ramps to the selected value over the first second of rotation, then remains constant, including during holding. The target is 45 degrees with a ±5-degree tolerance. Angular velocity and contact with both hands are also checked over the final two seconds.

We separately required the robot to remain within 15 degrees of facing forward. All four conditions missed this orientation criterion. Passing rotation and holding does not mean the entire task passed; posture control still needs improvement.

Match collision geometry to the visible shape

The original collision ring used rectangular pieces arranged in a circle, with up to about 5.8 mm cross-sectional deviation from the round appearance. We replaced it with an approximation made of 32 capsule shapes. The ring mass is 0.5 kg and contact friction coefficient is 0.5. These are test settings, not specifications of a particular real product.

An initial model that switched resistance according to speed showed numerical oscillation at low speed. Those results were excluded; the table contains only the corrected constant-load tests.

Measure slip as well as contact

We computed relative velocity at finger/ring contact points. Under 0.6 Nm, the median slip-speed indicator was 11.0 mm/s during rotation and 20.2 mm/s in the final two seconds. This includes small oscillations; it does not mean one point continuously slid by that distance. Persistent contact does not imply zero slip.

The contact-force-derived torque around the axis averaged 0.60 Nm over the same rotation interval. We aggregate contact points with normal force above 0.2 N over seconds 1–10 after rotation starts. Recording load, rotation and contact together reveals behavior not evident from video alone.

Constant opposing load does not reproduce real valve stiction, fluid pressure or seal friction. The arms use scripted control and the lower body uses Unitree’s public pretrained policy. No new ACT or reinforcement learning training was performed.

Next steps toward physical verification

The plan at the time of this experiment was to adapt the simulation to the conditions of a robot expected to be borrowed. These steps are future work, not part of the reported validation. Simulation success does not guarantee physical success; it helps define test conditions.

1. Match the robot and valve: identify the G1 hand configuration, valve dimensions, mounting height and starting/running torque. Replace the provisional 0–0.6 Nm settings with measurements.

2. Improve slip and posture: move from prescribed hand trajectories toward control that responds to contact and force, resolving the current slip and facing-direction failures.

3. Vary conditions: test valve placement, friction, resistance and control delay to establish the operating range and conditions for stopping. Align controller inputs with information available on the physical robot.

4. Compare in small physical tests: contact, holding and small-angle rotation on a test valve, then use measured differences to revise the model.

The plan is not to retrain whole-body control from the outset. First refine physical conditions and contact control, and consider whole-body learning if existing posture control proves insufficient. TeamZ builds equipment simulations and evaluates robot contact and motion.

Technology: Unitree G1 / Dex3, NVIDIA Isaac Sim 5.1.0, Isaac Lab 2.3.0, Blender, scripted IK and contact measurement.