day3_4_exp1 - MoE-ACT Model

Multi-skill Action Chunking Transformer with Mixture-of-Experts for pick_hanger task.

Model Info

• Architecture: MoE-ACT (10 experts, top-2 routing)
• Dataset: 26-03-04-pick_hanger_V30 (100 episodes)
• Training Steps: 100,010
• Robot: Piper dual-arm (7D + 7D = 14D state/action)
• Cameras: fisheye_left, fisheye_right, wide_mid, wide_top (256x256)
• FPS: 30

Checkpoints

Checkpoint	Steps	Description
060000/	60,000	Mid-training checkpoint
080000/	80,000	Late-training checkpoint
100010/	100,010	Final checkpoint

Normalization

normalization/26-03-04-pick_hanger_V30_norm.py contains pre-computed normalization constants:

• STATE_MEAN, STATE_STD (14D)
• ACTION_MEAN, ACTION_STD (14D)
• IMAGE_MEAN, IMAGE_STD (3D, ImageNet standard)

Usage

Download model

pip install huggingface_hub
huggingface-cli download QRP123/day3_4_exp1 --repo-type model --local-dir ./day3_4_exp1

Load model

from lerobot.policies.moe_act.modeling_moe_act import MoEACTPolicy

policy = MoEACTPolicy.from_pretrained("QRP123/day3_4_exp1/checkpoints/100010/pretrained_model")
policy.eval()

Load normalization

import sys
sys.path.append("./day3_4_exp1/normalization")
from normalization_26_03_04_pick_hanger_V30_norm import STATE_MEAN, STATE_STD, ACTION_MEAN, ACTION_STD

Training Config

• chunk_size: 100
• n_action_steps: 100
• batch_size: 8
• num_experts: 10
• num_experts_per_token: 2
• moe_balance_weight: 0.01