moonbot pi0 v2 — three blocks stack
pi0 fine-tuned from lerobot/pi0_base on
gdiazsrl/three_blocks_stack_sep22
(296 episodes; 263 train / 33 held out for validation, stratified over 9 tasks).
To run on the robot, see deploy/README.md — tokenizer,
contract and a script that makes a checkpoint load with no Hugging Face access.
Checkpoints
checkpoint-XXXXXX/ every 1,250 steps (80k samples) up to 16,250 (2 epochs).
Latest uploaded: checkpoint-007500 — training in progress.
Each is self-contained for the robot: weights, pre/post-processors, PaliGemma tokenizer,
rosetta contract, prepare_deploy.py, test_offline.py and a DEPLOY.md with the exact
commands. The final one also has training_state/.
Complete checkpoints (with optimizer state) for every published step are also on
the team OneDrive under DATA/Processed/pi0_v2_b64/.
Training (run v2)
| batch size | 64 |
| steps | 16,250 (= 2.0 epochs) |
| LR | cosine, peak 5e-5 -> 5e-6, warmup 250, decay over the full run |
| dtype / grad ckpt | bfloat16 / on (required: OOM without it on 96 GB) |
| validation | 10% of episodes per task, eval every 625 steps |
| GPU | RTX PRO 6000 Blackwell 96 GB, ~48 GB used |
Differences from run v1 (drashutoshspace/moonbot_pi0_three_blocks_stack):
batch 16 -> 64 with LR 2.5e-5 -> 5e-5 (same throughput, measured); LR schedule now
spans the whole run (v1's preset reached its floor by step ~20k of 20k);
validation split added; 2 epochs instead of 0.62.
Deployment contract
21-dim observation.state = joint_read (8) + tip_pos (7) + F_ee (6), 3 cameras,
8-dim action. Use deploy/rosetta_contract.yaml.