RoboColosseum/yam-g0.5

Galaxea G0.5 — MolmoAct2-BimanualYAM Full

• Base: OpenGalaxea/G05 g05-base (Qwen2.5-2B VLM + action expert), GalaxeaVLA repo, YAM recipe
• Data: Full MolmoAct2-BimanualYAM (32,246 episodes / 76M frames); flat 14-dim action/state [left_joint_0..5, left_gripper, right_joint_0..5, right_gripper]; cameras head_rgb=top, left_wrist_rgb, right_wrist_rgb
• Trained: full fine-tune, vision LR multiplier 0.1, DDP across 8x H200, 0.48 epochs (step 95,000 of ~197,917 steps/epoch)
• Optim: 8-bit Adam, LR 3.75e-5 (linear-scaled from 2.5e-5 base for GBS 384), cosine (min ratio 0.1), warmup 500, wd 0.03, 48/GPU x 8 GPUs → GBS 384
• Loss: 5.18 → 2.57 (cross-entropy over ActionCodec VQ tokens, codebook size 4096)
• Checkpoint: step 95,000 (final)
• Files: model.pt (step 95,000), dataset_stats.json, action_tokenizer.pt, task/data configs yam_full.yaml
• WandB: g05-yam/dmlvtlqz