siruku6/pi05_trial1500

π0.5 LIBERO — full-parameter fine-tune ladder (1,500 steps, then +2,500 steps)

Two full-parameter (all 4.14B params, nothing frozen) fine-tunes of π0.5 for LIBERO manipulation, continuing from siruku6/pi05_full_runpod: 001500 unfreezes the vision/language tower and adds camera-viewpoint augmentation for 1,500 steps, and ladder2500/002500 continues training the same run for another 2,500 steps with no other change.

Outcome. 001500 (this run's own checkpoint) was adopted and submitted — it is the weight behind the qualifier-round submission and the best PARC2026-finals Track1 score in the project (0.492). ladder2500/002500, produced by extending training on the same data and settings, was also adopted and is the final Track2 submission. However, the extra 2,500 steps mostly did not move the training loss (it plateaued after ~1,500 steps of full-parameter training — see Training setup below), and the intended camera augmentation effect did not show up in a 72-episode holdout eval (sign test p = 0.061, no condition improved). The score gains between checkpoints in this ladder track inference-configuration changes (EMA, resampling, adaptive step cadence) at least as much as the extra training.

Lineage

Stage	Weights	Steps	Trainable	Note
Base	lerobot/pi05_libero_base rev a217bfd3	—	—	Upstream π0.5 checkpoint
1	siruku6/pi05_stage1_24000	24,000	action expert only (693M / 4.14B)	JAIST H100, batch 16
2	siruku6/pi05_full_runpod	3,000	all (4.14B)	RunPod A100 80GB, batch 64
3 (this repo, 001500/)	pi05_trial1500	+1,500 (4,500 cumulative)	all (4.14B)	RunPod A100 80GB, batch 64, + camera augmentation
4 (this repo, ladder2500/002500)	pi05_ladder2500	+2,500 (7,000 cumulative)	all (4.14B)	Blackwell RTX PRO 6000, batch 64, same data/config as stage 3

Stage 3 and stage 4 differ in exactly one training-relevant variable: the step count. (Data loader parallelism and eval batch size also differ, but neither changes what is learned.) Note the optimizer scheduler's warmup (1,000 steps) is re-drawn from zero at the start of stage 4, since --init-from only loads weights, not optimizer/scheduler state — so stage 4 is not simply "stage 3 run longer" in effective learning rate terms.

Contents

001500/
  config.json
  model.safetensors                         # 9,354,050,752 bytes, sha256 52d3d0c7…
  policy_preprocessor.json
  policy_preprocessor_step_2_normalizer_processor.safetensors
  policy_postprocessor.json
  policy_postprocessor_step_0_unnormalizer_processor.safetensors
  train_config.json
  vlm/                                      # tokenizer files
ladder2500/
  002500/
    config.json
    model.safetensors                       # 9,354,050,752 bytes, sha256 cea403c9…
    policy_preprocessor.json
    policy_preprocessor_step_2_normalizer_processor.safetensors
    policy_postprocessor.json
    policy_postprocessor_step_0_unnormalizer_processor.safetensors
    train_config.json
    vlm/
  pi05_ladder2500.log                       # full training log
pi05_trial1500.log                          # full training log

(Intermediate checkpoints 000500/001000 of the trial1500 run, and 001000/001500/002000 of the ladder2500 run, were produced during training but are not kept in this repo.)

Training setup

	001500 (trial1500)	ladder2500/002500
Starting weights	pi05_full_runpod	pi05_trial1500/001500
Steps	1,500	2,500
Trainable params	4,143,404,816 / 4,143,404,816 (100%, nothing frozen)	same
Batch size	64	64
Optimizer	AdamW, betas (0.9, 0.95), eps 1e-8, weight decay 0.01, grad clip norm 1.0	same
LR schedule	cosine decay w/ warmup, peak 5e-6, decay 2.5e-6, warmup 1,000 steps, decay 30,000 steps	same schedule, warmup re-drawn from step 0
Precision	bfloat16	bfloat16
Seed	42	42
Image augmentation	on — brightness/contrast/saturation/hue jitter, sharpness jitter, affine, random-resized crop, perspective (max 4 of 8 transforms applied per sample)	same
Normalization	fit on this run's data (not pinned to a parent checkpoint)	same
Data	lerobot/libero_plus rev f3f49f42, 14,347 episodes	same
Hardware	RunPod, 1× A100 80GB	Blackwell (RTX PRO 6000 Blackwell Server Edition, 95.0 GB)
Wall-clock	4h 16m 05s	4h 48m 23s (6.85 s/step)
Framework	LeRobot v0.6.0	LeRobot v0.6.0
Final loss (10-step trailing avg)	0.2407 (from 0.2532 at step 500)	0.2311 at step 1,500 (best); 0.2331 at step 2,500 — did not improve monotonically past step 1,500

Evaluation

PARC2026-finals scores (Track1 / Track2, mean over 8 episodes each; higher is better). Same inference config across all rows below unless noted (num_inference_steps=10, n_action_steps=16, EMA α=0.3, samples=1) so the weight is the only variable, except where the inference config itself was also part of the change (noted inline):

Submission	Weight	Track1	Track2
Distributed baseline	π0.5 + LoRA, 5,000 steps	0.286	0.165
4.2.0 (pi05_full_runpod, this ladder's own starting point)	pi05_full_runpod	0.370	0.154
5.0.1	001500 (this repo)	0.492	0.233
8.0.0	ladder2500/002500 (this repo), same inference config as 5.0.1	0.475	0.261
8.1.0	ladder2500/002500, samples 1→3	0.410	0.258
8.2.0 (final Track2 submission)	ladder2500/002500 (this repo), + adaptive step-cadence (16→8 on high-variance steps)	0.324	0.318

A separate 72-episode holdout comparison of 001500 against its own starting point (pi05_full_runpod), isolating only the weight, found the targeted camera-augmentation conditions did not clearly improve (3 wins/2 losses on one held-out camera condition, 3/3 on another) and one un-targeted condition (large initial-pose perturbation) trended worse (0/5). Overall sign test across all held-out conditions: 9 improved / 20 regressed, p = 0.061 — not conclusive, but no clear net gain from the augmentation was found in that holdout.

Use

hf download siruku6/pi05_trial1500 --local-dir ./pi05_trial1500

from lerobot.policies.pi0.modeling_pi0 import PI0Policy  # π0.5 head is loaded the same way

policy = PI0Policy.from_pretrained("./pi05_trial1500/001500")
# or: PI0Policy.from_pretrained("./pi05_trial1500/ladder2500/002500")

Intended use and limitations

• Evaluated only in simulation (LIBERO tasks, simulated Franka Panda arm via robosuite/MuJoCo). No real-robot validation has been done.
• The camera-viewpoint augmentation added in this ladder was not shown to reliably help on held-out camera conditions (see Evaluation above); treat any single-run score gain over pi05_full_runpod with that caveat.
• Continuing training past ~1,500 full-parameter steps on this data/config showed diminishing-to-flat loss improvement; the ladder2500/002500 checkpoint's better Track2 score in this project came together with an inference-side change (adaptive step cadence), not from loss improvement alone.

License

These weights are a Model Derivative of Gemma (via PaliGemma inside π0.5) and are released under the Gemma Terms of Use. Use is also subject to the Gemma Prohibited Use Policy. See NOTICE for the third-party attributions that come with the base model, the data and the training code.