XRollout is a passionate community pushing the boundaries of robot learning. From VLA to Diffusion Policy, we pursue extreme generalization — one model that adapts to any robot, any task, any environment.
from xrollout import VLAConfig, Trainer
from xrollout.models import DiffusionPolicy
# 配置 VLA 模型
config = VLAConfig(
vision_encoder="siglip-so400m",
llm="qwen2-7b",
action_dim=7,
horizon=16
)
# 初始化训练器
trainer = Trainer(
config=config,
dataset="xrollout/robotics-v1",
policy=DiffusionPolicy(config)
)
# 开始训练
trainer.train(
batch_size=64,
num_epochs=100,
mixed_precision=True
)端到端视觉-语言-动作模型，支持 Diffusion Policy 动作解码。适用于机械臂操作、移动机器人导航等任务。
轻量级世界模型，支持3秒未来轨迹预测与不确定性估计。专为实时机器人避障设计。
基于 Diffusion 的轨迹规划器，生成平滑类人动作。支持动作分块与时间一致性约束。
大规模机器人操作数据集，包含 100k+ 条轨迹，涵盖抓取、放置、推拉等操作。支持多模态输入（RGBD、点云、本体感知）。
机器人安全基准测试数据集，包含 50k+ 危险场景，用于评估避障、碰撞检测、紧急制动等安全能力。
We believe the future of robotics isn't about training one model for one robot—
it's about extreme generalization.
We open-source not just code, but datasets, models, and training recipes. The community drives innovation faster than any single company.
One model should handle diverse robots, tasks, and environments. We pursue zero-shot and few-shot capabilities as the ultimate metric.
We optimize relentlessly—model architecture, training efficiency, inference speed. Every millisecond and every parameter matters.
We're hackers, researchers, and dreamers united by pure interest. No corporate mandates—just the joy of building the future.