PanoDP: Learning Collision-Free Navigation with Panoramic Depth and Differentiable Physics

2026-03-08

AI Summary

This work proposes a communication-free, safety-aware autonomous navigation method for complex environments characterized by partial observability, static structures, and dynamic obstacles. By integrating 360-degree panoramic depth perception with differentiable physics simulation, the approach constructs dense loss functions that encode collision avoidance and motion feasibility, replacing conventional sparse rewards to significantly enhance training stability and policy generalization. To the best of our knowledge, this is the first study to jointly leverage omnidirectional depth inputs and differentiable physical constraints for navigation learning. The method is validated on both a multi-agent ring-to-center benchmark and external distribution tests in AirSim. Experiments demonstrate substantial improvements in collision-free rate and task success across varying obstacle densities and dynamic behaviors compared to single-view and non-physics-guided baselines, with ablation studies confirming the effective utilization of panoramic information.

Technology Category

• Computer Vision: Low Level & Physics-based Vision
• Search and Optimization: Sampling/Simulation-based Search
• Intelligent Robots: Multimodal Perception & Sensor Fusion

Application Category

• Security and Privacy: Large-scale security measurements
• Responsible Web: Machine-in-the-loop, human agency and autonomy
• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications

Abstract

Autonomous collision-free navigation in cluttered environments requires safe decision-making under partial observability with both static structure and dynamic obstacles. We present \textbf{PanoDP}, a communication-free learning framework that combines four-view panoramic depth perception with differentiable-physics-based training signals. PanoDP encodes panoramic depth using a lightweight CNN and optimizes policies with dense differentiable collision and motion-feasibility terms, improving training stability beyond sparse terminal collisions. We evaluate PanoDP on a controlled ring-to-center benchmark with systematic sweeps over agent count, obstacle density/layout, and dynamic behaviors, and further test out-of-distribution generalization in an external simulator (e.g., AirSim). Across settings, PanoDP increases collision-free and completion rates over single-view and non-physics-guided baselines under matched training budgets, and ablations (view masking, rotation augmentation) confirm the policy leverages 360-degree information. Code will be open source upon acceptance.

Problem

• collision-free navigation
• partial observability
• static structure
• dynamic obstacles
• autonomous navigation

Innovation

• panoramic depth
• differentiable physics
• collision-free navigation
• multi-view perception
• reinforcement learning