piper_place_block_old0_old1_new2_bs8_50k

Policy trained for Piper randomized pick-and-place.

Experiment

Camera setup:

• cam0 = old0 / old eye_side
• cam1 = old1 / old eye_front
• cam2 = new2 / successful allnew cam_wrist

Purpose:

This model tests whether replacing only the old wrist camera old2 with the successful new2 wrist camera recovers the original old-camera failure.

Training

• policy: diffusion
• batch size: 8
• steps: 50000
• dataset: 3a279/piper_place_block_old0_old1_new2_random500

Artifacts

Training logs and loss plots are stored in:

training_artifacts/