NavThinker: Action-Conditioned World Models for Coupled Prediction and Planning in Social Navigation

Abstract

Social navigation requires robots to act safely in dynamic human environments. Effective behavior demands thinking ahead: reasoning about how the scene and pedestrians evolve under different robot actions rather than reacting to current observations alone. This creates a coupled prediction-planning challenge, where robot actions and human motion mutually influence each other. To address this challenge, we propose NavThinker, a future-aware framework that couples an action-conditioned world model with on-policy reinforcement learning. The world model operates in the Depth Anything V2 patch feature space and performs autoregressive prediction of future scene geometry and human motion; multi-head decoders then produce future depth maps and human trajectories, yielding a future-aware state aligned with traversability and interaction risk. Crucially, we train the policy with DD-PPO while injecting world-model think-ahead signals via: (i) action-conditioned future features fused into the current observation embedding and (ii) social reward shaping from predicted human trajectories. Experiments on single- and multi-robot Social-HM3D show state-of-the-art navigation success, with zero-shot transfer to Social-MP3D and real-world deployment on a Unitree Go2, validating generalization and practical applicability. Webpage: https://github.com/hutslib/NavThinker.

Figures (5)

• Figure 1: NavThinker: future-aware social navigation. Social navigation requires the robot to reach a goal in environments shared with dynamic pedestrians. Red regions highlight interaction zones where the robot and pedestrians may conflict. Top: egocentric depth observations at two timesteps. Bottom: the world model imagines action-conditioned futures for the corresponding observations. By imagining how the scene evolves under different candidate actions, NavThinker anticipates potential social conflicts and selects safe, socially compliant actions.
• Figure 2: Overview of the NavThinker framework. Our framework consists of two modules: a world model that learns action-conditioned scene dynamics (top), and an imagination-augmented planner policy trained with DD-PPO (bottom). During World Model Learning, depth observations, actions, and robot states are extracted from the Replay Buffer. A frozen DA-V2 ViT encoder extracts patch embeddings, and a Causal Attention Transformer autoregressively predicts future latent features. A Depth Decoder and a Human Trajectory Decoder are trained alongside a latent consistency loss to anchor representations to scene geometry and human motion. During Policy Learning, the Observation Encoder produces the current embedding from depth and robot states. The Imagination module queries the world model under each candidate action, generating action-conditioned future features that are fused with the current embedding via Feature Fusion. The fused representation feeds into the DD-PPO Actor-Critic network. Predicted human trajectories additionally provide Reward Shaping, coupling prediction with planning in an imagine-then-act loop.
• Figure 3: World model imagination quality. Left: current depth observation with RGB, third-person view, and top-down map for reference. Right: after executing the action at the previous timestep, we show (from top to bottom) the ground-truth depth, the world model's predicted depth, the PCA visualization of ground-truth DA-V2 features, and the PCA visualization of predicted features. The world model produces faithful depth predictions and latent features that closely match the ground truth across different actions.
• Figure 4: Qualitative comparisons on Social-HM3D. Navthinker produces safe actions under different social situations. Rows 1–2 (Blind corner): A* collides when turning, while ours avoids the pedestrian. Rows 3–4 (Intersection): Falcon collides at the intersection, while ours safely yields and passes. Rows 5–6 (Front approach): Habitat-official collides in a head-on encounter, while ours gives way for the pedestrian. Red boxes mark baseline failures; green checkmarks indicate safe behaviors.
• Figure 5: Real-world deployment on a Unitree Go2 robot. In this scenario, the robot navigates around a group of people and reaches its goal while maintaining safe distances.