Ratatouille: Imitation Learning Ingredients for Real-world Social Robot Navigation

2025-09-21

AI Summary

To address collision-prone navigation and insufficient safety in real-world social environments, this paper proposes Ratatouille—a novel offline imitation learning framework that eliminates online trial-and-error risks of reinforcement learning and enables zero-shot deployment on physical robots. Methodologically, it integrates behavioral cloning with a lightweight spatiotemporal perception network and a robust data augmentation strategy, trained exclusively on 11 hours of real-world campus navigation data—without additional annotations or simulation-based pretraining. Experiments demonstrate that Ratatouille reduces collisions per meter by 6× and improves task success rate by 3× over standard behavioral cloning. It further validates safety and generalization across dense campus corridors and public food courts. The core contribution is the first demonstration of highly robust social navigation achieved solely from offline expert demonstrations—without online fine-tuning—thereby significantly enhancing safety and reliability for real-world robotic deployment.

Technology Category

• Machine Learning: Imitation Learning & Inverse Reinforcement Learning
• Natural Language Processing: Safety and Robustness
• Intelligent Robots: Multimodal Perception & Sensor Fusion

Application Category

• Economics, Online Markets and Human Computation: Social networks and social learning
• Responsible Web: Human-perceived consequences of algorithmic deployment on the web
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

Abstract

Scaling Reinforcement Learning to in-the-wild social robot navigation is both data-intensive and unsafe, since policies must learn through direct interaction and inevitably encounter collisions. Offline Imitation learning (IL) avoids these risks by collecting expert demonstrations safely, training entirely offline, and deploying policies zero-shot. However, we find that naively applying Behaviour Cloning (BC) to social navigation is insufficient; achieving strong performance requires careful architectural and training choices. We present Ratatouille, a pipeline and model architecture that, without changing the data, reduces collisions per meter by 6 times and improves success rate by 3 times compared to naive BC. We validate our approach in both simulation and the real world, where we collected over 11 hours of data on a dense university campus. We further demonstrate qualitative results in a public food court. Our findings highlight that thoughtful IL design, rather than additional data, can substantially improve safety and reliability in real-world social navigation. Video: https://youtu.be/tOdLTXsaYLQ. Code will be released after acceptance.

Problem

• Scaling reinforcement learning for real-world social robot navigation safely
• Improving imitation learning performance for collision-free robot navigation
• Enhancing safety and reliability in dense social environments

Innovation

• Pipeline and model architecture improving imitation learning
• Reduces collisions by 6x and triples success rate
• Validated with 11+ hours of real-world campus data