Model Card for Hume-Libero_Spatial

• Paper: https://arxiv.org/abs/2505.21432
• Homepage: https://hume-vla.github.io
• Codebase: 🦾 Hume: A Dual-System VLA with System2 Thinking

  

Optimal TTS Args

Uses

• If you want to reproduce the results in paper, follow the instruction
• If you want to directly use the model:
  python

  1from hume import HumePolicy 2import numpy as np 3 4# load policy 5hume = HumePolicy.from_pretrained("/path/to/checkpoints") 6 7# config Test-Time Computing args 8hume.init_infer( 9 infer_cfg=dict( 10 replan_steps=8, 11 s2_replan_steps=16, 12 s2_candidates_num=5, 13 noise_temp_lower_bound=1.0, 14 noise_temp_upper_bound=1.2, 15 time_temp_lower_bound=1.0, 16 time_temp_upper_bound=1.0, 17 post_process_action=True, 18 device="cuda", 19 ) 20) 21 22# prepare observations 23observation = { 24 "observation.images.image": np.zeros((1,224,224,3), dtype = np.uint8), # (B, H, W, C) 25 "observation.images.wrist_image": np.zeros((1,224,224,3), dtype = np.uint8), # (B, H, W, C) 26 "observation.state": np.zeros((1, 7)), # (B, state_dim) 27 "task": ["Lift the papper"], 28} 29 30# Infer the action 31action = hume.infer(observation) # (B, action_dim)

Training and Evaluation Details

Citation