TrossenRoboticsCommunity/pi05-block-transfer-bs8

pi0.5 — Block Transfer (batch size 8)

Fine-tuned pi0.5 vision-language-action policy for the block transfer task on the Trossen AI Stationary bimanual platform.

Model details

Architecture	pi0.5 (PaliGemma gemma_2b backbone + gemma_300m action expert)
Base model	lerobot/pi05_base
Framework	LeRobot
Robot platform	Trossen AI Stationary (bimanual, 14-DoF)
Cameras	cam_high, cam_low, cam_left_wrist, cam_right_wrist (480×640 RGB)
Action chunk	50 steps, 10 inference steps
Precision	bfloat16
License	Apache 2.0

Training

Hyperparameter	Value
Batch size	8
Training steps	100,000
Learning rate	2e-5 (cosine decay to 2.5e-6 over 30k steps, 1k warmup)
Optimizer	AdamW (β=(0.9, 0.95), wd=0.01, grad clip 1.0)
Gradient checkpointing	Enabled
Hardware	1× NVIDIA H200 (141 GB VRAM) via Trossen Cloud Service
Dataset	TrossenRoboticsCommunity/trossen_ai_stationary_block_transfer

Usage

Load the policy with LeRobot:

from lerobot.policies.pi0_5.modeling_pi0_5 import PI05Policy

policy = PI05Policy.from_pretrained(
    "TrossenRoboticsCommunity/pi05-block-transfer-bs8"
)

See the LeRobot pi0.5 docs for the full inference / rollout setup on real hardware.

Files

File	Purpose
config.json	Policy configuration
model.safetensors	Model weights (~7.5 GB)
policy_preprocessor.json + *_normalizer_processor.safetensors	Input normalizer (state/action quantiles)
policy_postprocessor.json + *_unnormalizer_processor.safetensors	Output unnormalizer

Related models

• TrossenRoboticsCommunity/pi05-block-transfer-bs8 — this model
• TrossenRoboticsCommunity/pi05-block-transfer-bs32 — same task, batch size 32