ロボット基盤モデルがシミュレーションでは高性能なのに、カメラの位置や照明が変わった途端に失敗する原因、実は「見た目のズル」だったかもしれません。 タイトル: Breaking the Vision-Action Shortcut: Latent Interface Training for Generalizable Robotics Foundation Models URL: 📝 概要 訓練分布内でたまたま行動と相関しているだけの無関係な視覚手がかりに頼ってしまう「視覚-行動ショートカット」を断ち切る学習手法「Latent Interface Training（LIT）」を提案しています。 ❗ 解決する課題 ロボット基盤モデルは分布内では強いのに、視覚的な分布シフト（カメラ視点・照明・センサノイズなど）が起きると性能が大きく落ち込む問題がありました。 ⚙️ 方法論 まず画像を使わずポーズ目標だけで行動部分を事前学習し、次に100個の潜在トークンを介して視覚情報を必ず経由させ、ポーズ再構成損失で監督する2段階構成になっています。 🤖 ユースケース π0.5・MolmoAct2・FAST-WAM・ImageWAMという4種類の異なるロボット基盤モデルすべてに適用でき、フレームワークを選ばない汎用的な手法です。 📊 実験結果 LIBERO-Plusの分布外評価で全モデルが3.87〜10.70ポイント改善。実世界のロボット操作でも照明変化時に+16.7ポイント、特定タスクでは+60ポイントという大きな改善を達成しています。 見た目に頼らず本質的な空間情報で判断させる発想が、実運用ロボットの信頼性向上に直結しそうです。 #ロボティクス# #VLA#

# Learning Palantir Foundry 🚀 "How many screens do I need to open just to understand one customer?" Object Views answer that pain by bundling everything about a single object into one screen. 📌 Title and Feature URL Title: オブジェクトビュー URL: 📝 Overview Object Views act as the central hub for everything related to a specific object. They consolidate properties, linked objects, metrics and analytics, dashboards, and operational applications into a single unified interface. For example, an Airport object view can integrate flight timelines, delay-handling workflows, and location data in one place. In practice, an Object View becomes the daily "home screen" that frontline users open to start their work. 🔧 How It Works Object Views are highly configurable by builders: - They support multiple formats and sizes, so appearance and interaction patterns can be tailored to the task. - They combine properties (attributes), linked objects, metrics, analytics, and dashboards into one display. - They can be embedded throughout the platform wherever the object appears. - Configuration happens in the Ontology Manager under the "Object views" tab, and version tabs at the top let you switch between format variations. - Selecting "Edit views" opens the configuration editor or the underlying Workshop module. - The system also supports version management, panel variations, commenting, and Marketplace product integration. 🛠 Practical Usage - In Ontology Manager, select the target object type and use the "Object views" tab to preview and configure. - Lay out core information, related objects, operation history, and embedded dashboards so everything the team needs is on one screen. - Beyond viewing, embed action types so users can trigger status changes or assignments directly from the view. - Create multiple formats to show different layouts per role (for example, sales view vs. maintenance view). 🎯 Use Cases - Customer 360: one launchpad combining transaction history, inquiries, related orders, and account owners for sales. - Equipment record: a maintenance home screen with sensor values, service history, related parts, and open tickets. - Case management: a single view of stakeholders, due dates, approval status, and next actions on a case object. ⚠️ Caveats - Views depend on the quality of the underlying ontology modeling (object types and link types); a weak foundation limits view quality. - Overloading a view confuses users, so design role-specific layouts that show only what each role needs. - Editing requires appropriate permissions to the Ontology Manager and the underlying Workshop module. #PalantirFoundry# #DataPlatform#