moonbot pi0 — three blocks stack, WITHOUT force/torque
pi0 fine-tuned from lerobot/pi0_base on
gdiazsrl/lerobot_sep23_no_tf.
Comparison run. Counterpart of drashutoshspace/moonbot_pi0_three_blocks_stack
(trained WITH force/torque). Training config is identical to that run (verified by diffing the
parsed config against its saved train_config.json), and the dataset is identical frame by frame
(same episodes, timestamps, tasks, actions; 50/50 videos byte-identical) except that the 6 F_ee
wrench dimensions are removed from the state. Any behavioural difference is due to force/torque input.
Deployment contract: 15-dim state = joint_read (8) + tip_pos (7), NO F_ee force/torque; 3 cameras; 8-dim action.
Checkpoints
checkpoint-005000, -010000, -015000, -020000 (same steps as the counterpart run).
Each folder is self-contained for the robot: weights, pre/post-processors, PaliGemma tokenizer,
this run's rosetta contract, prepare_deploy.py, test_offline.py and DEPLOY.md.
The final checkpoint also includes training_state/. Complete checkpoints with optimizer state are
also on the team OneDrive under DATA/Processed/pi0_no_tf/.
Training
| batch size | 16 |
| steps | 20,000 |
| optimizer / schedule | pi0 preset: AdamW lr 2.5e-5, betas 0.9/0.95, clip 1.0; cosine warmup 1000 -> 2.5e-6 |
| dtype / grad ckpt | bfloat16 / on |
| seed | 1000 (default) |