Robot arm stuttering during action execution with fine-tuned GR00T model

Description: Hi team, I’ve fine-tuned a GR00T model and deployed it on our robotic arm (SO100/SO101). While the model is able to produce roughly correct actions, the execution is not smooth — the robot exhibits noticeable stuttering or jerky motions between action segments. It looks like the actions are being executed in discrete chunks rather than as a continuous motion.

Details:

• Robot: SO101

• GR00T version: 4ea96a16b15cfdbbd787b6b4f519a12687281330

• Deployment environment: CPU i9-14900K GPU Nvidia 4090

Observed behavior:

• The arm pauses or slightly jitters between motion chunks.

• Movement is not fluid, even when the target poses are simple or nearby.

• Reducing model temperature doesn’t seem to help significantly.

Questions:

1. Is this kind of stuttering expected due to GR00T’s chunked architecture?

2. Are there recommended techniques to smooth the output or post-process actions for more continuous control?

3. Should we consider fine-tuning with a loss or reward specifically targeting motion smoothness?

Any guidance or insights would be appreciated. Thanks!

below is my eval figure in testset.