Learning Social Navigation from Positive and Negative Demonstrations and Rule-Based Specifications

TL;DR

The paper addresses safe, adaptive social navigation for mobile robots in crowded human environments by integrating data-driven rewards with explicit safety specifications. It introduces PioneeR, a framework that learns a density-based reward from positive and negative demonstrations and augments it with rule-based obstacle avoidance and goal-reaching; a lookahead teacher evaluates short-horizon rollouts and a distillation step yields a real-time, uncertainty-aware student policy. The density reward leverages an RKHS formulation with a smooth leveraged kernel, while the teacher–student pipeline delivers both adaptability and safety, plus risk signals via MDN-based uncertainty. Empirical results in synthetic static scenes, elevator co-boarding simulations, and real-world demonstrations show higher success rates and efficiency than baselines, with predictive uncertainty offering actionable safety cues. This work provides a practical pathway to socially aware robot navigation that balances data-driven flexibility with explicit safety guarantees in dynamic human environments.

Abstract

Mobile robot navigation in dynamic human environments requires policies that balance adaptability to diverse behaviors with compliance to safety constraints. We hypothesize that integrating data-driven rewards with rule-based objectives enables navigation policies to achieve a more effective balance of adaptability and safety. To this end, we develop a framework that learns a density-based reward from positive and negative demonstrations and augments it with rule-based objectives for obstacle avoidance and goal reaching. A sampling-based lookahead controller produces supervisory actions that are both safe and adaptive, which are subsequently distilled into a compact student policy suitable for real-time operation with uncertainty estimates. Experiments in synthetic and elevator co-boarding simulations show consistent gains in success rate and time efficiency over baselines, and real-world demonstrations with human participants confirm the practicality of deployment. A video illustrating this work can be found on our project page https://chanwookim971024.github.io/PioneeR/.

Figures (4)

• Figure 1: Overview of the proposed framework. A. Reward learning: (a) density-based reward maps are constructed from positive and negative demonstrations, and (b) augmented with rule-based specifications for obstacle avoidance and goal reaching. B. Teacher policy: short-horizon candidate rollouts are simulated, scored with the combined reward, and used to select safe and adaptive supervisory actions. C. Student policy: the teacher’s guidance is distilled into a compact policy conditioned on LiDAR observations, enabling real-world deployment.
• Figure 2: Reward maps and resulting trajectories with synthetic dataset. (a) Learning-based reward map trained with positive samples, highlighting both feasible corridors. (b) Learning-based reward map trained with positive and negative samples, reducing reward near humans. (c) Reward map combining learning-based and rule-based components, yielding trajectories with greater clearance.
• Figure 3: Simulation snapshots of elevator co-boarding scenarios. (A) HR-RL: Human on the right, Robot on the left. (B) HL-RR: Human on the left, Robot on the right.
• Figure 4: Real-world elevator co-boarding experiments with human participants. (A) Single policy across diverse scenarios. (B) Scenario involving multiple human interactions.