Hi all,
I’m working on training the Interbotix RX200 manipulator to perform a task where it grabs a cube and lifts it to a specified position. This is essentially the same task as the Isaac-Lift-Cube-Franka-v0 environment, which trains the Panda Franka robot.
I have Replaced the Franka Panda arm with the Interbotix RX200 manipulator, reduced the size of the cube so it properly fits in our gripper and we have adjusted the range of coordinates where the cube and the target frame are generated.
After training, the RX200 tends to “faceplant” near the cube rather than successfully grasping it.From analyzing the first 200 training epochs, we see that it is attempting to grab the cube, but it fails consistently. We have tried to fix this by experimenting with different learning rates, different reward and penalty weights but had no significant improvement in performance.
The Questions are:
- Could this failure stem from a mistake in our environment setup (e.g., incorrect parameters, robot-specific dynamics, etc.)?
- Is it more likely that we are encountering a local maximum during training, preventing the manipulator from learning an effective grasping strategy?
Any insights, suggestions, or similar experiences would be greatly appreciated. We’re especially curious about debugging techniques for custom environments or ways to refine the training process for non-standard manipulators.