pi05_combined_cont4k

A pi0.5 checkpoint fine-tuned on the local/libero_combined_bowl5_initpose dataset (see siruku6/libero_combined_bowl5_initpose), continuing full-parameter training from siruku6/pi05_combined_initpose_full (checkpoint 005500, itself fine-tuned from all parameters starting at lerobot/pi05_libero_base, revision a217bfd3b14673cf2ce597e69997ab21866438dd).

Training

Starting point	siruku6/pi05_combined_initpose_full checkpoint 005500
Additional steps	4,000
Batch size	64
Trainable parameters	all 4.14B (vision encoder and VLM unfrozen)
Optimizer	AdamW, lr 5e-6, weight decay 0.01, betas (0.9, 0.95), grad clip norm 1.0
Image augmentation	disabled
Action chunk	chunk_size=50, n_action_steps=10
Dataset	local/libero_combined_bowl5_initpose (111 tasks, 21,642 episodes)

The full configuration as actually used is in 004000/train_config.json.

Evaluation

Measured on the same Track2 validation protocol as three other candidates (84 rows x 5 episodes, EGL rendering; see the dataset card of libero_combined_bowl5_initpose for the data-side comparison):

	Track2 validation total score	Success rate	Collision rate	LIBERO90-half total score	Weighted score
This checkpoint	0.3801	0.755	0.162	0.1194	0.3280
pi05_bowl5_full (selected candidate, trained on libero_plus_bowl5 instead)	0.4055	0.798	0.138	0.0841	0.3412

The gap to the selected candidate (-0.013 in weighted score) is within the project's "no meaningful difference" margin (0.02) but this checkpoint was not the one carried forward, mainly by tie-break ordering rather than a statistically clear loss (paired sign test p=0.063 on 420 matched trials). It scored significantly above C (pi05_combined_initpose_full, p=0.026), confirming that continuing training partially recovered ground lost from the libero_combined_bowl5_initpose data choice, without fully closing the gap to libero_plus_bowl5-trained training. This repository is published for transparency and reproducibility of that comparison.

Modifications to the base model

model.safetensors is modified relative to the base checkpoint: it is the result of gradient updates performed by the author. No other file carried over from the base model is modified.

License and redistribution

This is a Model Derivative of Gemma (via PaliGemma inside pi0.5) and is therefore distributed under, and subject to, the Gemma Terms of Use.

Gemma is provided under and subject to the Gemma Terms of Use found at ai.google.dev/gemma/terms

By using or further distributing these weights (or any derivative of them) you agree to:

• the Gemma Terms of Use, a copy of which governs this repository and must be passed on to any recipient you redistribute to; and
• the Gemma Prohibited Use Policy, whose use restrictions apply to this derivative and to every derivative made from it.

Reference implementation of pi0.5: openpi (Apache License 2.0). Training data is derived from LIBERO (Lifelong-Robot-Learning/LIBERO, MIT) and LIBERO-plus (Sylvest/LIBERO-plus, MIT).

Provided as-is, without warranty of any kind.