Using G1 in a field-based reasoning experiment, the results actually relied on memory rather than reasoning. Really need help😭 #147
Description
Our company appreciates the training, deployment, and inference process services provided by your company.
We are using the G1 humanoid wheeled robot.
Currently, we have reached the inference and deployment stage, using a fully configured training solution (70G+).
**Our scenario is simple: the initially open gripper slowly moves to a beverage bottle, then clamps the bottle, and ends.
Inference has revealed the following: In scenarios with a bottle, the robot performs the bottle-clamping action flawlessly 10 times; 2 times the bottle enters the gripper's mouth but is not clamped; and 1 time the arm moves slowly and erratically. (We haven't yet considered subtly moving the bottle's position because the robot doesn't track it.) The strangest and most bizarre thing is: In scenarios without a bottle, the robot correctly clamps the bottle 6 times out of 10 times, even when doing so against thin air. We believe that clamping should not be performed when there is no bottle.**
We collected 50 data points, trained for 28 hours using a cloud card, and used 3 checkpoints. The above results represent our best checkpoint performance.
We look forward to guidance from experienced developers if we are fortunate enough to receive such assistance. Thank you!
