Abstract
We present Scene Graph Contrastive (SGC) learning, a novel auxiliary task for training embodied agents. By aligning an agent's internal latent "belief" state with a dynamically updated scene graph through contrastive learning, agents develop a structured semantic understanding of their environment. This approach yields significant performance gains in Object Navigation tasks—up to 12% in success rate—without requiring additional sensors or ground-truth maps during inference. Our results demonstrate that incorporating structured history via scene graphs allows agents to navigate more efficiently and reduce false-positive goal identifications.
Introduction
Training AI agents for embodied navigation is a fundamental challenge in robotics. Standard reinforcement learning (RL) often suffers from sparse terminal rewards, requiring millions of steps to learn basic behaviors. While auxiliary tasks like depth estimation or frame prediction can accelerate learning, they often focus on low-level pixel data and fail to capture the high-level semantic context necessary for complex tasks, such as understanding that "bread is typically found in a kitchen."
Existing attempts to use structured scene graphs—which represent objects and their relationships—often rely on complex graph decoders that are computationally expensive and difficult to optimize. SGC learning addresses these issues by using a simple yet effective contrastive objective to align agent beliefs with structured environmental history.
Methodology
Iterative Scene Graph Construction
As the agent explores an environment, the system builds a ground-truth scene graph. Nodes in the graph represent the agent, rooms, and objects, while edges represent spatial and semantic relationships such as Sees, On, Contains, or Above.
Crucially, this graph is persistent; it maintains a history of the episode. Even when an object leaves the agent's field of view, it remains in the graph. This persistence teaches the agent to "remember" previously seen landmarks and their relative locations.
Scene Graph Contrastive (SGC) Loss
The core of our method is the SGC loss. The agent maintains an internal "belief" state ($b_t$) using a CLIP-ResNet50 visual encoder and a GRU recurrent unit. Simultaneously, the ground-truth scene graph is encoded into a vector ($g_t$) using a Graph Attention Network.
Instead of attempting to reconstruct the entire graph from the belief state, we use contrastive learning to maximize the similarity between $b_t$ and its corresponding $g_t$ in a shared embedding space. This forces the agent's latent representation to capture the structural and semantic information contained within the graph. Importantly, the scene graph is only required during training; at test time, the agent operates using only standard RGB input.
Experimental Results
We evaluated SGC across several benchmarks, including ProcTHOR, RoboTHOR, and MultiON. The results show consistent and substantial improvements over pure RL baselines:
- Object Navigation Performance: SGC achieved a 12% absolute gain in Success Rate (SR) on RoboTHOR and a 9% gain on AI2-iTHOR.
- Multi-Object Navigation: In the MultiON-3 task (finding three objects in sequence), SGC improved the success rate from 13.3% to 21.4%.
- Reliability and Stopping: SGC-trained agents were significantly more accurate at identifying when they had reached a goal. In RoboTHOR, standard RL agents incorrectly signaled success in 45% of episodes, whereas SGC agents reduced this false-positive rate to 14.9%.
- Generalization: Linear probing analysis confirmed that SGC-trained beliefs are more effective at predicting object visibility and historical presence, even for novel object categories not seen during specific task training.
Conclusion
Scene Graph Contrastive learning provides a powerful yet computationally efficient way to inject structured semantic knowledge into embodied agents. By prioritizing structure over raw pixels and maintaining a persistent history of environmental relationships, SGC enables agents to learn more robust representations for navigation. Future work will explore more sophisticated graph encoders and the use of learned, rather than heuristic, relationship definitions.
Summary updated