You can deploy this model and start earning money today!

ASethi04/
MolmoAct2-BimanualYAM-oranges-12k

Name: MolmoAct2-BimanualYAM-oranges-12k
Availability: InStock

lerobot

safetensors

robotics

vision-language-action

molmoact2

bimanual

brandonyang/yam-vive-teleop

allenai/MolmoAct2-BimanualYAM

finetune

apache-2.0

Views

No views yet

MolmoAct2-BimanualYAM — "Put all oranges in the bowl" (12k steps)

Fine-tune of allenai/MolmoAct2-BimanualYAM on brandonyang/yam-vive-teleop (80 teleop episodes, 74,927 frames, bimanual YAM, 30 fps) for "Put all oranges in the bowl".

Companion to ASethi04/MolmoAct2-BimanualYAM-oranges (5,000 steps). This model trains 2.4× longer and is expected to be the stronger policy.

What changed vs the 5k run

	5k run	this model (12k)
Steps / epochs	5,000 / 4.3	12,000 / 10.3
Action-expert LR	1.5e-4	1e-4
LoRA adaptation vs reference recipe	60%	145%
Expert:LoRA travel ratio	3.0	2.0 (matches reference)
Saturation jitter	0.5–1.5	0.8–1.2
Hue jitter	±0.05	±0.02

The VLM LoRA learning rate is hardcoded at 5e-5 in the training code, so adaptation on that path scales only with step count. 12,000 steps restores it to 145% of the reference recipe while keeping the expert-to-LoRA ratio at the reference's 2.0. The narrower color jitter reflects that this task is defined by locating orange objects.


Trainable params	727,296,544 / 5,591,928,304 (13%)
VLM	LoRA r=64, α=16, dropout=0.05 @ LR 5e-5
Action expert	fully fine-tuned @ LR 1e-4
Global batch	64 (8 GPUs × 8)
Optimizer	AdamW β=(0.9,0.95), ε=1e-6, wd=0, clip 1.0
Schedule	cosine, 600-step warmup, decay ratio 0.1
Precision	bfloat16 + gradient checkpointing
Action mode	`both` (discrete FAST + flow matching), 8 flow timesteps
Chunk / executed	30 / 30 (1 s @ 30 Hz)
Action space	14-D absolute joint pose
Cameras	`observation.images.{top,left,right}` @ 480×270
Normalization	quantile q01/q99; grippers raw
Split / seed	100/0 (all 80 episodes) / 1000

Final training step:

 step:12K smpl:768K ep:820 epch:10.25 loss:0.665 grdn:2.058 lr:5.0e-06 updt_s:2.410 data_s:0.074 smp/s:26 mem_gb:25.86 discrete_ce_loss:0.661 discrete_z_loss:0.000 action_flow_loss:0.003

Usage notes

Set inference_action_mode="continuous" — the saved config has None.
Do not pass norm_tag.
Task string must match training exactly: Put all oranges in the bowl.
Consider n_action_steps < 30 at deployment: executing all 30 open-loop means replanning at 1 Hz, which limits reactivity on a sequential multi-object task.

bash

lerobot-policy-server \
  --policy.pretrained_name_or_path=ASethi04/MolmoAct2-BimanualYAM-oranges-12k \
  --policy.inference_action_mode=continuous \
  --policy.model_dtype=bfloat16 --policy.device=cuda --host=0.0.0.0 --port=8081

Limitations

Single task, 80 demonstrations, no held-out validation set. Validate on hardware with a no-motion action probe before arming the robot.