pravsels/build_block_tower_baseline_6mix_joints_only

pi0.5 Build Block Tower 6-Mix — Joints-Only Ablation

Fine-tuned pi0.5 checkpoint for the block-tower task, trained with loss restricted to the first 7 joint dimensions only. EEF channels of state/action are zeroed and excluded from the flow-matching loss via action_dim_mask.

Experiment

• Objective: Test whether removing EEF supervision improves joint-space policy quality on the block-tower task.
• Weight init: weights/pi05_base/params (pi0.5 base weights).
• Total steps: 50,000 (completed)
• Final loss: 0.0112 (step 49,900)

Config

• Config name: pi05_build_block_tower_baseline_6mix_joints_only
• Model: pi0.5 (pi05=True, action_horizon=50)
• Batch size: 36
• Learning rate: 5e-5 cosine decay (1k warmup)
• Optimizer: AdamW (gradient clip norm 1.0)
• EMA decay: 0.999
• Delta actions: enabled
• Action space: 17D canonical (first 7 joint dims active, remaining 10 EEF dims masked)
• joints_only: True — action_dim_mask=[True]*7+[False]*10
• Norm stats: reused from pi05_build_block_tower_baseline_6mix/retain/step_49999/alpha_0.5/assets

Dataset

6 HuggingFace datasets: villekuosmanen/build_block_tower plus dAgger_build_block_tower_1.0.0 through 1.4.0 (340 episodes total).

Checkpoint Hashes

Verify integrity with:

cd checkpoints/<step> && find params -type f | sort | xargs sha256sum | sha256sum

Step	Loss	SHA-256
49,999	0.0112	7cd730580e57c6ae390ff2e0fc41491941b6b4d10e0e5cf959e6243034eabb45

W&B

• Training dashboard

Repo Structure

assets/                      # Norm stats, valid_indices.txt
checkpoints/49999/params/    # Model weights (params only)
README.md                    # This file
TRAINING_LOG.md              # Training log

Usage

from openpi.training.config import get_config
from openpi.serving.policy_server import PolicyServer

config = get_config("pi05_build_block_tower_baseline_6mix_joints_only")
server = PolicyServer(config, checkpoint_path="checkpoints/49999/params")