X-VLA LIBERO: Duplicated Steps in Environment and Policy PreProcessor during Eval

[gianlucageraci]

Ticket Type

🐛 Bug Report (Something isn't working)

Environment & System Info

Description

Note that this issue coincides heavily with #2954.

I identified a redundancy in the preprocessing pipeline when evaluating the XVLA policy within the LIBERO environment.

The pipeline instantiates two processor chains:

1. Environment Preprocessor: via make_env_pre_post_processors → make_xvla_libero_pre_post_processors
2. Policy Preprocessor: via make_pre_post_processors → make_xvla_pre_post_processors

Both chains currently append XVLAImageNetNormalizeProcessorStep and XVLAAddDomainIdProcessorStep. Because the pipeline executes Env Preprocessor → Policy Preprocessor, these transformations are applied twice to the input observations.

Reproduction / Code Analysis

The duplication occurs in lerobot/policies/xvla/processor_xvla.py.

1. Environment-Specific Processor (make_xvla_libero_pre_post_processors)
Adds normalization and domain ID:

def make_xvla_libero_pre_post_processors() -> tuple[...]:
    # ...
    pre_processor_steps.extend(
        [LiberoProcessorStep(), XVLAImageNetNormalizeProcessorStep(), XVLAAddDomainIdProcessorStep()]
    )
    # ...

2. General Policy Processor (make_xvla_pre_post_processors) Also adds normalization and domain ID:

def make_xvla_pre_post_processors(...) -> tuple[...]:
    # ...
    input_steps = [
        # ...
        XVLAImageToFloatProcessorStep(),
        XVLAImageNetNormalizeProcessorStep(), # DUPLICATE
        XVLAAddDomainIdProcessorStep(),       # DUPLICATE
        # ...
    ]

This leads to double normalization which in turns leads to erroneous input values and redundant processing.

Suggested Fix

Remove the general XVLA steps from the environment-specific processor (make_xvla_libero_pre_post_processors). It should likely only handle LiberoProcessorStep and leave the standard normalization to the policy processor.

def make_xvla_libero_pre_post_processors() -> tuple[...]:
    # ...
    # Removed XVLAImageNetNormalizeProcessorStep and XVLAAddDomainIdProcessorStep
    pre_processor_steps.extend(
        [LiberoProcessorStep()]
    )
    # ...

Additional Remarks

Currently this issue is mitigated when directly evaluating the pre-trained lerobot/xvla-libero checkpoint available on huggingface as the checkpoint there does not include the XVLAImageToFloatProcessorStep and the XVLAImageNetNormalizeProcessorStep (which I think is incorrect, see #2954).
When I first finetuned lerobot/xvla-libero on LIBERO with the additional XVLAImageToFloatProcessorStep and the XVLAImageNetNormalizeProcessorStep steps in the preprocessing pipeline and then evaluated this model I ran into the double normalization error which led to negative input values.

Context & Reproduction

No response

Relevant logs or stack trace

Checklist

• I have searched existing tickets to ensure this isn't a duplicate.
• I am using the latest version of the main branch.
• I have verified this is not an environment-specific problem.

Additional Info / Workarounds

No response