drashutoshspace/moonbot_pi0_three_blocks_stack

moonbot pi0 — three blocks stack, WITH force/torque

pi0 fine-tuned from lerobot/pi0_base (PaliGemma 2B + 300M action expert; full fine-tune: vision encoder and VLM trainable), trained on gdiazsrl/three_blocks_stack_sep22 (296 episodes, 519,684 frames at 30 Hz, 9 language instructions; MoonBot 7-DOF arm, block stacking).

Input/output: 21-dim state = joint_read (8: 7 joints + gripper) + tip_pos (7: end-effector pose) + F_ee (6: force/torque); 3 RGB cameras (front, eef, right); 8-dim action (7 joint positions + gripper).

Comparison run. One of six policies comparing pi0 / SmolVLA / ACT with vs without force/torque, all trained with batch size 16, 20,000 steps, seed 1000 and no validation split; otherwise each uses its own LeRobot preset. The two datasets are the same episodes and frames, except that the 6 F_ee dimensions are removed from the state in the no-F/T one. The counterpart of this model without force/torque is moonbot_pi0_no_tf.

model	F/T	repo	W&B run
pi0	with	moonbot_pi0_three_blocks_stack	bkbszs91
pi0	without	moonbot_pi0_no_tf	ahb0i6l9
SmolVLA	with	moonbot_smolvla_tf	rn1je6y9
SmolVLA	without	moonbot_smolvla_no_tf	9i3olvax
ACT	with	moonbot_act_tf	uw4alfzk
ACT	without	moonbot_act_no_tf	a7hpuoto

Training

initialization	lerobot/pi0_base
parameters	4.03B (all trainable)
language input	yes
action chunk	50 (all 50 executed)
optimizer	AdamW, lr 2.5e-5, betas (0.9, 0.95), weight decay 0.01, grad clip 1.0
LR schedule	cosine with 1,000 warm-up steps, decaying towards 2.5e-6 over 30,000 steps (stops at 20,000)
precision	bfloat16, gradient checkpointing
batch size / steps	16 / 20,000 (about 0.62 epochs)
seed	1000
GPU	1x NVIDIA RTX PRO 6000 Blackwell (96 GB)
training time	9.9 h
final training loss	0.0145 (this policy's own objective; not comparable across algorithms)
framework	LeRobot 0.6.1
training curves	https://wandb.ai/drmishra-space/lerobot/runs/bkbszs91

Checkpoints

checkpoint-005000, checkpoint-010000, checkpoint-015000, checkpoint-020000 (final). Each checkpoint-XXXXXX/ folder holds the weights (model.safetensors), config.json, train_config.json and the pre/post-processor files. checkpoint-020000/ also includes training_state/ (optimizer, scheduler, RNG, step) for resuming. This repo has no deployment bundle: deploy with the 21-dim rosetta contract moonbot1.yaml from the rosetta fork. Complete checkpoints, including the optimizer state, are also on the team OneDrive under DATA/Processed/pi0_checkpoint_020000 (final checkpoint).

Deployment

Per Gustavo: open config.json in the downloaded checkpoint and remove invalid fields such as pretrained_revision before loading with rosetta/policy_bridge_node.py.

hf download drashutoshspace/moonbot_pi0_three_blocks_stack --include "checkpoint-020000/*" --local-dir model
python src/rosetta/rosetta/policy_bridge_node.py --ros-args \
  -p contract_path:=<rosetta>/moonbot1.yaml \
  -p policy_path:=$PWD/model/checkpoint-020000