Downloads · 30 days
22
45% of all-time downloads
ASethi04/MolmoAct2-BimanualYAM-oranges-ultrawide-wristonly-12k
MolmoAct2-BimanualYAM-oranges-ultrawide-wristonly-12k is a robotics model from ASethi04. Use it for the robotics task on the model card, and read the license before you ship it in a product. It is set up for lerobot. The card lists the license as apache-2.0.
Fine-tune of allenai/MolmoAct2-BimanualYAM on brandonyang/yam-ultrawide-teleop (75 episodes, 108,726 frames, 30 fps — ultrawide wrist cameras at 312×416, top at 270×480).
Downloads · 30 days
22
45% of all-time downloads
All-time downloads
49
Public
Parameters
5.6B
11.5 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors11.5 GB · 100%
How the weights are stored.
BF165.4B · 97%
From the Hugging Face model README
Fine-tune of allenai/MolmoAct2-BimanualYAM on brandonyang/yam-ultrawide-teleop (75 episodes, 108,726 frames, 30 fps — ultrawide wrist cameras at 312×416, top at 270×480).
In-the-wild variant: NO top camera. Trained with image_keys=[left, right] only — the model consumes two ultrawide wrist views and nothing else. Deploy with exactly these two cameras; do not supply a top image.
Trained with the lerobot PR#4249 fixes (fp32 flow targets, DDP-correct loss normalization,
episode-boundary supervision, llm_residual_dropout=0.1). inference_action_mode=continuous
is baked into the config.
| Steps / epochs | 12,000 / ~7.1 |
| Global batch | 64 (8 GPUs × 8) |
| VLM | LoRA r=64 @ 5e-5; action expert full FT @ 1e-4 |
| Schedule | cosine, 600 warmup, decay ratio 0.1 |
| Augmentation | brightness/contrast/saturation 0.8–1.2, hue ±0.02, sharpness 0.5–1.5, affine ±5°/5% |
| Action space | 14-D absolute joint pose, chunk 30/30 @ 30 Hz |
| Normalization | quantile q01/q99; grippers raw |
| Split / seed | 100/0 / 1000 |
Final training step: step:12K smpl:768K ep:530 epch:7.06 loss:1.150 grdn:1.400 lr:5.0e-06 updt_s:1.707 data_s:0.052 smp/s:36 mem_gb:23.05 discrete_ce_loss:1.146 discrete_z_loss:0.000 action_flow_loss:0.004
norm_tag.Put all oranges in the bowl.--robot.max_gripper_delta=0.05 (data needs 0.05/tick).Single task, 75 demos, no held-out val. Validate with a no-motion action probe before arming.