Hume-vla/Libero-Goal-2

Model Card for Hume-Libero_Goal

A Dual-System Visual-Language-Action model with System-2 thinking trained on Libero-Goal.

• 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

s2_candidates_num=5
noise_temp_lower_bound=1.0
noise_temp_upper_bound=1.0
time_temp_lower_bound=0.9
time_temp_upper_bound=1.0

Uses

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

  from hume import HumePolicy
  import numpy as np
  
  # load policy
  hume = HumePolicy.from_pretrained("/path/to/checkpoints")
  
  # config Test-Time Computing args
  hume.init_infer(
      infer_cfg=dict(
          replan_steps=8,
          s2_replan_steps=16,
          s2_candidates_num=5,
          noise_temp_lower_bound=1.0,
          noise_temp_upper_bound=1.0,
          time_temp_lower_bound=0.9,
          time_temp_upper_bound=1.0,
          post_process_action=True,
          device="cuda",
      )
  )
  
  # prepare observations
  observation = {
      "observation.images.image": np.zeros((1,224,224,3), dtype = np.uint8), # (B, H, W, C)
      "observation.images.wrist_image": np.zeros((1,224,224,3), dtype = np.uint8), # (B, H, W, C)
      "observation.state": np.zeros((1, 7)), # (B, state_dim)
      "task": ["Lift the papper"],
  }
  
  # Infer the action
  action = hume.infer(observation) # (B, action_dim)
  

Training and Evaluation Details

# source ckpts
2025-05-02/08-10-44_libero_goal_ck8-16-1_sh-4_gpu8_lr5e-5_1e-5_1e-5_2e-5_bs16_s1600k/0090000
# original logs
2025-06-12/20-53-09+08-10-44_libero_goal_ck8-16-1_sh-4_gpu8_lr5e-5_1e-5_1e-5_2e-5_bs16_s1600k_0090000_s1-8_s2-16_s2cand-5_ntl-1.0_ntu-1.0_ttl-0.9_ttu-1.0.log

Citation

@article{song2025hume,
  title={Hume: Introducing System-2 Thinking in Visual-Language-Action Model},
  author={Anonimous Authors},
  journal={arXiv preprint arXiv:2505.21432},
  year={2025}
}