siruku6/pi05_camdrop2500

π0.5 LIBERO — scene-camera-dropout variant (2,500 steps)

A full-parameter fine-tune of lerobot/pi05_libero_base on lerobot/libero_plus, trained with the third-person ("scene") camera blanked to zeros with probability 0.5 per sample, so that the policy is pushed to rely on the wrist camera.

This is a published negative result. In our simulated LIBERO evaluation the dropout variant did not beat its identically-trained non-dropout sibling, and we stopped the line of work. The weights and the training log are here so the comparison is reproducible.

Lineage

Stage	Weights	Steps	Note
Base	lerobot/pi05_libero_base @ a217bfd3b14673cf2ce597e69997ab21866438dd	—	π0.5, 4.14 B params
+1	siruku6/pi05_full_runpod	3,000	full-parameter fine-tune, no dropout
+2	siruku6/pi05_trial1500 001500	1,500	no dropout
+3 (this repo)	000500 … 002500	2,500	scene-camera dropout p=0.5

The only difference from the non-dropout sibling run (same starting weights, same 2,500 steps, same hyper-parameters, same seed) is the dropout.

Contents

Five checkpoints, saved every 500 steps. Each directory is the flattened content of a LeRobot pretrained_model/: config.json, model.safetensors (9,354,050,752 bytes), train_config.json. logs/pi05_camdrop2500.log is the full training log.

000500/  001000/  001500/  002000/  002500/  logs/

Training setup

Objective / policy	π0.5 (pi05), flow-matching action expert, action chunk 50, n_action_steps 10
Trainable	all 4,143,404,816 parameters (freeze_vision_encoder=false, train_expert_only=false)
Batch / LR	64 / 5e-6 peak, cosine decay with warmup
Precision	bfloat16, gradient checkpointing on
Seed	42
Image augmentation	brightness / contrast / saturation / hue / sharpness / affine / resized-crop / perspective, up to 4 per sample
Data	lerobot/libero_plus — 14,347 episodes, 2,238,036 frames, 40 tasks, 20 fps, Franka Panda
Hardware / time	1× NVIDIA RTX PRO 6000 Blackwell, 4 h 46 min 37 s, 6.84 s/step, 49.4 GB VRAM
Framework	LeRobot v0.6.0

The dropout

For each training sample the scene-camera image is replaced with zeros with probability 0.5; the wrist camera is never touched. The realised rate, counted every 20,000 frames, was 0.500 / 0.502 / 0.502 / 0.505.

It costs nothing to run. Against the non-dropout sibling: step time 6.833–6.835 s vs 6.832–6.834 s, data time 0.007 s in both, 49.42 GB VRAM in both.

It does change what the model learns. Training loss sits above the sibling's throughout, and the gap stops closing after ~900 steps:

Steps	1–100	401–500	801–900	1401–1500	2001–2100	2401–2500
dropout	0.2662	0.2566	0.2558	0.2506	0.2500	0.2532
no dropout	0.2382	0.2354	0.2362	0.2304	0.2304	0.2354
gap	+11.8%	+9.0%	+8.3%	+8.8%	+8.5%	+7.6%

Read this as: what the scene camera was carrying is genuinely gone, and 2,500 steps of wrist-only practice did not make it back.

Use

hf download siruku6/pi05_camdrop2500 --include '002500/*' --local-dir ./camdrop2500

Then point LeRobot at the local directory:

lerobot-train --policy.type=pi05 --policy.pretrained_path=./camdrop2500/002500 ...

Note that the pretrained_model/ level is flattened away in this repo, so lerobot-train --resume against the Hub path will not find train_config.json where it expects it. Downloading first and passing a local --policy.pretrained_path works.

Intended use and limitations

Research artifact. Trained and evaluated only in LIBERO simulation with a simulated Franka Panda — there is no real-robot validation, and nothing here should be run on physical hardware without your own safety review. Performance outside the LIBERO task and camera setup is unknown.

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 dataset and the training code.