pravsels/pi05-arx5-multitask-micro-advantaged

pi0.5 ARX5 Multitask Micro Advantaged

Fine-tuned pi0.5 checkpoint for multi-task manipulation with ARX5 arms, trained on a 14-dataset micro mix with advantaged valid-index filtering.

Experiment

• Objective: Fine-tune PI0.5 on the micro training mix with advantaged valid indices; compare to baseline variant.
• Weight init: weights/pi05_base/params (pi0.5 base weights).
• Total steps: 30,000 (completed)
• Final loss: 0.0080 (step 29,900)

Config

• Config name: pi05_arx5_multitask_micro_advantaged
• Model: pi0.5 (pi05=True, action_horizon=50)
• Batch size: 36
• Learning rate: 5e-5 cosine decay (1k warmup, decay over 100k steps)
• Optimizer: AdamW (gradient clip norm 1.0)
• EMA decay: 0.999
• Delta actions: enabled (delta joints, absolute grippers)
• Per-timestep action normalization: enabled (auto from delta actions)
• Action space: 14D bimanual (single-arm 7D padded to 14D with loss masking)

Dataset

14 LeRobot datasets from training_mix_micro.json (all villekuosmanen/* repos). Filtered by valid_indices.txt (advantaged indices).

Checkpoint Hashes

Verify integrity with:

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

Step	Loss	SHA-256
25,000	0.0089	1648c67a7ac44d377f28f316384bdcab72af4422237f9f9485e1e77a02c6a65c
29,999	0.0080	aff337d89dd426388303855ed8fca784f5b5615b33cbad14f26dfbe8688caa88

W&B

• Training dashboard

Repo Structure

assets/                      # Norm stats, valid_indices.txt, training_mix_micro.json
checkpoints/<step>/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_arx5_multitask_micro_advantaged")
server = PolicyServer(config, checkpoint_path="checkpoints/<step>/params")