YunzeLiu/dreamzero-e2

DreamZero — Epoch 2 checkpoint

LoRA fine-tune of DreamZero-AgiBot on the RoboDojo arx_x5 dataset (6-DOF dual-arm).

• Checkpoint step: 17,500 (epoch 2)
• Action: relative for arm/head/waist, absolute for grippers
• Base model: DreamZero-AgiBot (7-DOF AgiBot G1 pretrain)

Training loss

Red dashed lines mark epoch boundaries (step 8,750 and 17,500).

step	total	arm	ee (gripper)	aux
100	0.418	0.427	1.037	0.026
500	0.288	0.255	0.979	0.016
1,000	0.238	0.199	0.513	0.005
2,000	0.200	0.158	0.179	0.011
3,500 (warmup end)	0.169	0.201	0.175	0.002
5,000	0.140	0.137	0.124	0.002
8,750 (epoch 1)	0.109	0.154	0.178	0.011
12,000	0.098	0.057	0.097	0.004
17,500 (epoch 2)	0.118	0.051	0.033	0.001

Summary (step 0 → 17,500):

• Total loss: 0.42 → 0.12 (−72%)
• Arm loss: 0.43 → 0.05 (−88%)
• Gripper loss: 1.04 → 0.03 (−97%)
• Aux loss (head/waist/velocity, all zeros for this embodiment): 0.03 → 0.001 (live-channel mask working)

Files

File	Purpose
model.safetensors	LoRA weights (317 MB)
experiment_cfg/conf.yaml	Full Hydra training config (required at inference)
config.json	HF Trainer config
trainer_state.json	Per-step loss history
loss_curves.png	Loss visualization

Bug fixes vs original DreamZero training

Includes the fixes shipped in Luminis-Sim/XPolicyLab#31:

1. J3-locked joint routing for 6-DOF arms onto AgiBot G1's 7-DOF slot layout
2. Mask-aware action loss (live-channel detection)
3. Gripper normalization q99 → min_max
4. Per-modality (arm/ee/aux) loss logging