Rohan Paul@rohanpaul_ai

2026-07-12 04:44·24分钟前

AI 摘要

LingBot-VLA 2.0是一个全身体控制机器人策略，通过55维动作格式统一控制手臂、夹爪、灵巧手、头部、腰部和移动底座。该模型在20种机器人配置上训练，将9万小时原始数据过滤为5万小时高质量真实机器人数据。采用稀疏MoE模块，每个动作token使用专用网络，共享专家保留通用技能。额外训练信号要求模型预测当前和未来深度及视频特征，以跟踪物体几何和场景变化。在Agilex GM-100上，LingBot-VLA 2.0达到66.2%进度和34.4%成功率，优于pi0.5的59.1%和32.2%。在长时域移动任务中，LingBot-VLA 2.0在域内和域外测试均领先pi0.5。

Most VLAs （Vision-Language-Action Models） handle task variety only inside narrow， fixed bodies；

LingBot-VLA 2.0 from @robbyant_brain trains one policy across 20 configurations with whole-body control.

Also to avoid the damaging noise of a robot datasets， LingBot-VLA 2.0 filters 90，000 raw hours into 50，000 cleaner high-quality real-robot data hours.

Supports 20 robot configurations and whole-body degree-of-freedom control.

So LingBot-VLA 2.0 from @robbyant_brain is a whole-body robot policy. It controls arms， grippers， dexterous hands， the head， waist， and mobile base through one 55-dimensional action format， so the same model can learn from very different robot bodies.

About 90，000 raw robot hours are checked for jerky motion， broken signals， camera mismatch， blur， dropped frames， and long static periods before 50，000 hours remain.

Human videos are filtered for hand-object interaction， then camera motion and hand pose are reconstructed as action data.

A sparse Mixture of Experts module lets each action token use a few specialized networks while a shared expert preserves common skills.

Another training signal asks the model to predict current and future depth plus video features， pushing it to track object geometry and likely scene changes during the next action chunk.

On Agilex GM-100， it reaches 66.2% progress and 34.4% success， ahead of pi0.5 at 59.1% and 32.2%. It also leads pi0.5 across both long-horizon mobile tasks in both in-domain and out-of-distribution tests， although several individual GM-100 tasks still favor other models.

🧵 1.

Rohan Paul@rohanpaul_ai · X

56导出 Markdown