pi05-base-mini-chair-openpi

Full fine-tune of Physical Intelligence pi0.5 (pi05_base) trained with openpi (PyTorch backend) on JonathanGiegold/mini-chair-base-put-only.

Files

• model.safetensors — fine-tuned weights (final step 12,963)
• metadata.pt — openpi checkpoint metadata
• assets/JonathanGiegold/mini-chair-base-put-only-v21/norm_stats.json — required state/action normalization stats for inference
• training_summary.png — training curves (loss / grad-norm / LR)

Optimizer state is intentionally omitted (inference/deploy only).

Training curves

Raw per-step loss shows spikes from flow-matching noise-level sampling; the smoothed trend falls 0.31 -> ~0.035 and is flat at the end. Gradient-norm spikes coincide with those steps but were absorbed by gradient clipping (norm 1.0, fired on ~2% of steps), so the optimizer stayed stable throughout.

Model

Architecture	pi0.5 (PaliGemma gemma_2b + action expert gemma_300m)
Action horizon	16
State / action dim	14 (padded to 32)
Precision	bfloat16
Cameras (3)	context, wrist_left, wrist_right

Data

Dataset	JonathanGiegold/mini-chair-base-put-only (LeRobot v2.1 on-box conversion)
Transforms	repack + delta joint actions (mask [Tx6, F, Tx6, F]), prompt_from_task
Normalization	quantile (norm_stats computed on this dataset)

Hyperparameters

Steps	12,964 (5 epochs)
Batch size	32
Num workers	8
Seed	42
EMA	disabled
Optimizer	AdamW
betas (b1, b2)	0.9, 0.95
eps	1e-8
weight_decay	1e-10
grad clip norm	1.0
LR schedule	cosine decay with warmup
warmup steps	1,000
peak LR	2.5e-5
decay steps	12,964
floor LR	2.5e-6

Run

Hardware	1x H100 80GB
Wall time	11h 29m (3.19 s/step)
VRAM	~41.6 GB
Final smoothed loss	~0.035