kuroll/motus_vla_arena

Motus VLA-Arena Training Hyperparameters

1. Common

Key	Value
action_dim	7
state_dim	8
num_video_frames	8
video_height	384
video_width	320
global_downsample_rate	1
video_action_freq_ratio	6

2. Dataset

Key	Value
type	lerobot
max_episodes	null
image_aug	false
task_mode	single
task_name	null
params.repo_id	vla_arena
params.embodiment_type	vla_arena
params.enable_t5_fallback	true
params.t5_folder_name	t5_embedding
params.t5_text_len	512

3. Model

Key	Value
wan.precision	bfloat16
vlm.precision	bfloat16
vlm.frozen	true
action_expert.hidden_size	1024
action_expert.ffn_dim_multiplier	4
action_expert.norm_eps	0.00001
und_expert.hidden_size	512
und_expert.ffn_dim_multiplier	4
und_expert.norm_eps	0.00001
und_expert.vlm.input_dim	2048
und_expert.vlm.projector_type	mlp3x_silu
time_distribution.timestep_sample_method	logit_normal
time_distribution.sigmoid_scale	1.0
time_distribution.min_t	0.0
time_distribution.max_t	1.0
inference.num_inference_timesteps	10
loss_weights.video_loss_weight	1.0
loss_weights.action_loss_weight	1.0
ema.enabled	false
ema.update_after_step	0
ema.inv_gamma	1.0
ema.power	0.75
ema.min_value	0.0
ema.max_value	0.9999

4. Training

Key	Value
batch_size	4
max_steps	45000
learning_rate	2.0e-5
weight_decay	0.01
scheduler_type	linear
warmup_steps	500
cycle_length	45000
f_max	0.99
f_min	0.4
grad_clip_norm	0.5
use_amp	true
find_unused_parameters	false

5. System

Key	Value
system.val_interval	1000
system.num_workers	16
system.pin_memory	true

6. Notes

• batch_size 为单卡 batch size；采用8卡h100，总 batch size 为 32。