[Bug] H0.5 inference server incompatible with released LIBERO/RoboCasa benchmark client

[MarkfuGod]

Hi @zawnpn

Description

Running the released Being-H0.5 inference server (BeingH.inference.run_server_vla) with the
benchmark client end-to-end fails before/at rollout due to three compatibility issues.

Bugs

1. get_modality_config endpoint takes a payload it shouldn't.
   In Being-H05/BeingH/inference/beingh_service.py, get_modality_config is registered with the
   default requires_input=True. The server's request handler therefore calls
   policy.get_modality_config(request["data"]), but BeingHPolicy.get_modality_config(self) takes
   no arguments → TypeError: get_modality_config() takes 1 positional argument but 2 were given.
2. BaseDataConfig is missing modality_config().
   BeingHPolicy.get_modality_config() returns self.data_config.modality_config(), but neither
   BaseDataConfig nor any subclass defines modality_config() → AttributeError. This blocks
   client-side modality discovery.
3. State observations are sent as float64.
   The LIBERO benchmark reaches a real rollout but fails in state preprocessing: simulator
   observations are float64, while the H0.5 transform pipeline (StateActionToTensor →
   StateActionTransform) keeps the numpy dtype (torch.from_numpy) and the model expects float32.
   RoboCasa uses the same observation path and is affected identically.

Affected files

• Being-H05/BeingH/inference/beingh_service.py
• Being-H05/configs/data_config.py
• Being-H05/BeingH/benchmark/utils/policy.py
• Being-H05/BeingH/benchmark/robocasa/run_robocasa_eval_fast.py
  Note: bugs (1) and (2) also exist in the tutorials/Being-H-EDU/ workspace (same
  beingh_service.py registration and same data_config.py without modality_config()).

Proposed fix

• Register the get_modality_config endpoint with requires_input=False.
• Add a generic BaseDataConfig.modality_config() returning the video/state/action/language
  ModalityConfigs from the existing *_KEYS and obs_indices/action_indices.
• Cast LIBERO and RoboCasa benchmark state observations to float32 before sending.

Validation

• py_compile of the four files; local server start; verified ping, get_modality_config,
  remote dummy get_action.
• LIBERO eval (robosuite 1.4.0): spatial 493/500, object 485/500, goal 490/500, long 475/500.
  Fix in Fix H0.5 benchmark inference compatibility #50.