Introducing Mobi-π: Mobilizing Your Robot Learning Policy. Our method: ✈️ enables flexible mobile skill chaining 🪶 without requiring additional policy training data 🏠 while scaling to unseen scenes 🧵↓

We introduce the "policy mobilization" problem. Given a manipulation policy trained on data collected from limited viewpoints, policy mobilization aims to find an optimal robot pose in an unseen environment to successfully execute this policy on a mobile platform. 1/6

What makes a good policy mobilization method? We propose a new method consisting of four simple steps: 1️⃣ Capture the test-time scene 2️⃣ Create a 3D Gaussian Splat 3️⃣ Compute hybrid scores of rendered views 4️⃣ Optimize! 2/6

We show in simulated and real-world experiments that our policy mobilization method is 💡 more data-efficient than an end-to-end mobile manipulation policy 💡 more robust than naive object navigation methods 3/6

We demonstrate that policy mobilization enables flexible deployment of locally-trained policies on a mobile platform in the real world, including: ✅ scaling to large scenes ✅ generalization to unseen room layouts and object positions ✅ composition of atomic skills 4/6

To effectively study policy mobilization, we develop the Mobi-π framework, which includes a simulated task suite, mobilization feasibility metrics, baselines, and visualization tools. 5/6

Want to learn more? Website: Code: arXiv: Team: @yjy0625, Isabella Huang, @brandontvu, Max Bajracharya, @contactrika, @leto__jean Collaboration between @ToyotaResearch and @StanfordIPRL 6/6

Great work! Congrats Jingyun and the team!!!👏

Thanks Yunfan!

Mobi-π's ability to scale to unseen scenes without additional training data is what excites me most! This could really push mobile robotics forward. ✈️

Love how Mobi-π ditches constant retraining—huge for real-world robotics! Can totally picture these skills in action in kitchens or hospitals. What’s the next space you’re targeting? 👏

@Rollup_News rollup