Downloads · 30 days
18
39% of all-time downloads
MolmoAct2-BimanualYAM-oranges-12k-fixed-nodrop
ASethi04/MolmoAct2-BimanualYAM-oranges-12k-fixed-nodrop
MolmoAct2-BimanualYAM-oranges-12k-fixed-nodrop 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-vive-teleop (80 teleop episodes, 74,927 frames, bimanual YAM, 30 fps).
All-time downloads
46
Public
Parameters
5.6B
11.5 GB on disk
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors11.5 GB · 100%
Parameter types
How the weights are stored.
BF165.4B · 97%
Base models
Datasets
Model card
From the Hugging Face model README
MolmoAct2-BimanualYAM — "Put all oranges in the bowl" (12k steps, PR#4249 fixes, no residual dropout)
Fine-tune of allenai/MolmoAct2-BimanualYAM on brandonyang/yam-vive-teleop (80 teleop episodes, 74,927 frames, bimanual YAM, 30 fps).
Sibling of ASethi04/MolmoAct2-BimanualYAM-oranges-12k-fixed:
identical in every respect except llm_residual_dropout=0 (that model uses the upstream
default 0.1). The pair is a controlled A/B on residual dropout for small-data LoRA
fine-tuning. Both carry the training fixes from
lerobot PR#4249: fp32 flow-matching
targets, distributed-correct loss normalization, supervised episode-boundary actions,
and correct embedding freezing.
Offline eval (final checkpoint, in-sample; camera-shuffle probe = vision dependence)
| Model | Joint L1 | Gripper L1 | Vision ratio |
|---|---|---|---|
| -oranges-12k (pre-fixes) | 0.0515 | 0.0036 | 1.53x |
| -oranges-12k-fixed (dropout 0.1) | 0.0528 | 0.0051 | 1.51x |
| this model (dropout 0) | 0.0519 | 0.0039 | 1.53x |
Training
| Trainable params | 727,296,544 / 5,591,928,304 (13%) |
| VLM | LoRA r=64, α=16, dropout=0.05 @ LR 5e-5 (train_mode_vlm=lora) |
| Action expert | fully fine-tuned @ LR 1e-4 |
| Steps / epochs | 12,000 / 10.25 · final loss 0.664 |
| Global batch | 64 (8 GPUs × 8) |
| Schedule | AdamW β(0.9,0.95), cosine, 600-step warmup |
| Precision | bfloat16 + gradient checkpointing (fp32 flow targets) |
| Action mode | both, chunk 30/30, 8 flow timesteps, 14-D absolute joint pose |
| Cameras | observation.images.{top,left,right} @ 480×270 |
| Augmentation | brightness/contrast/saturation 0.8–1.2, hue ±0.02, sharpness 0.5–1.5, affine ±5°/5% |
| Normalization | quantile q01/q99; grippers raw · seed 1000 · 100/0 split |
Usage notes
inference_action_mode="continuous"is baked into this checkpoint — no flag needed.- Do not pass
norm_tag— stats come from the fine-tuning dataset. - Task string must match training exactly:
Put all oranges in the bowl. - Client gripper rate limit:
--robot.max_gripper_delta=0.05. - All three cameras at 16:9.
Limitations
Single task, 80 demonstrations, no held-out validation. Validate on hardware with a no-motion action probe before arming the robot.
- Task
- Robotics
- Library
- lerobot
- License
- apache-2.0
- Created
- Aug 6, 2026
- Updated
- Aug 6, 2026