XVLA-Base-0.88B

License & Attribution

This repository contains weights or code derived from the XVLA foundational architecture developed by Hugging Face and the LeRobot/XVLA Authors.

• License: Distributed under the https://apache.org.
• Original Model Base: https://huggingface.co/lerobot/xvla-base
• Copyright Notice: Copyright 2025-2026 Hugging Face & The LeRobot/XVLA Authors.
• Research paper introducing this model: https://arxiv.org/pdf/2510.10274

This is XVLA-Base model cloned from Hugginface "lerobot/xvla-base" repository. This was createed for ready-to-use custom model for easy inference during Hackathon challenge.

Script to load model:

import numpy as np
from PIL import Image
import torch
import lerobot
from lerobot.policies.xvla.modeling_xvla import XVLAPolicy
from lerobot.policies.factory import make_pre_post_processors
from lerobot.datasets.lerobot_dataset import LeRobotDataset

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

model_id = "Man1103/XVLA-Base-0.88B"

policy = XVLAPolicy.from_pretrained(model_id).to(device)

print("XVLA model successfully loaded!")

Script to load model specific preprocessor and postprocessor assets:

from lerobot.policies.factory import make_pre_post_processors

preprocess, postprocess = make_pre_post_processors(
    policy.config,
    model_id,
    preprocessor_overrides={"device_processor": {"device": str(device)}},
)

print("XVLA pre-processor and post-processor successfully loaded!")

Script for sample inference:

from lerobot.datasets.lerobot_dataset import LeRobotDataset

# load a lerobotdataset (we will replace with a simpler dataset)
dataset = LeRobotDataset("lerobot/libero")

# pick an episode
episode_index = 0

# each episode corresponds to a contiguous range of frame indices
from_idx = dataset.meta.episodes["dataset_from_index"][episode_index]
to_idx   = dataset.meta.episodes["dataset_to_index"][episode_index]

# get a single frame from that episode (e.g. the first frame)
frame_index = from_idx
frame = dict(dataset[frame_index])

batch = preprocess(frame)
with torch.inference_mode():
    pred_action = policy.select_action(batch)
    # use your policy postprocess, this post process the action
    # for instance unnormalize the actions, detokenize it etc..
    pred_action = postprocess(pred_action)