Aether258/pi05_bi_vitac_clean_smash_24

pi05_bi_vitac — clean_smash_24 (Step 60k)

Checkpoint at step 60,000 / 250,000 of a Pi0.5 bimanual visual-tactile policy fine-tuned on 24 manipulation datasets covering clean and smash task families.

Model Architecture

Component	Details
Base model	Pi0.5 (pi05)
Vision-language backbone	PaliGemma (Gemma 2B)
Action expert	Gemma 300M
Tactile encoder	AnyTouch CLIP-B/16, 2-frame variant (full fine-tune)
AnyTouch pool tokens	14
State dim	20
Action dim	20
Action horizon	50

Visual observations pass through SigLIP; tactile observations pass through AnyTouch. The AnyTouch encoder is fully fine-tuned (LoRA rank = 0).

Training Details

Hyperparameter	Value
Total steps	250,000
This checkpoint	60,000
Batch size	128 (2 × 64, FSDP)
Optimizer	AdamW
Weight decay	1e-4
Gradient clip norm	1.0
LR schedule	Cosine decay
Warmup steps	5,000
Peak LR	3e-5
Decay steps	250,000
Final LR	6e-7
EMA decay	0.999
Save interval	10,000 steps
Base weights	gs://openpi-assets/checkpoints/pi05_base/params

Training Data

24 LeRobot datasets from EricChen06, covering two task families:

Clean (pick-and-place): green_clean_01–04, red_clean_01–04, blue_clean_01–04

Smash: white_smash_01/03/04/05, black_smash_01–04, yellow_smash_01–04

Checkpoint Contents

60000/
├── _CHECKPOINT_METADATA
├── assets/
├── params/          # EMA model parameters (for inference)
└── train_state/     # Full optimizer state (for resuming training)

Use the params/ sub-tree for inference and the train_state/ sub-tree only if resuming training.

Usage

This checkpoint is designed for use with ManiSkill-vitac and the openpi training framework.

# Inference
python deploy_scripts/infer.py \
  --checkpoint_path /path/to/60000 \
  --config pi05_bi_vitac

To resume training from this checkpoint, point weight_loader at the downloaded 60000/ directory.

Citation

If you use this model, please cite the underlying Pi0.5 and AnyTouch works.