Dejavu: Towards Experience Feedback Learning for Embodied Intelligence

Researchers introduce Dejavu, a post-deployment learning framework that enables frozen Vision-Language-Action policies to improve through experience retrieval and feedback networks. The system allows embodied AI agents to continuously learn from past trajectories without retraining, improving task performance across diverse robotic applications.

Dejavu addresses a critical bottleneck in embodied AI deployment: the inability of frozen models to adapt after entering real-world environments. Traditional approaches require complete retraining or fine-tuning to improve performance, which is computationally expensive and impractical for deployed systems. This research introduces an Experience Feedback Network that retrieves contextually relevant past experiences and uses them to condition action predictions, effectively creating a memory-augmented decision system without modifying the underlying policy weights.

The technical approach combines semantic similarity matching with reinforcement learning objectives, training the EFN to identify when historical trajectories are applicable to current situations. This positions the work within the broader trend of retrieval-augmented AI systems gaining traction across language models, vision systems, and robotics. The ability to learn continuously from deployment experiences represents a shift toward more adaptive, self-improving autonomous systems.

For the robotics and embodied AI industry, this framework has significant implications for scalability and real-world deployment viability. Agents that improve their performance over time without costly retraining cycles reduce operational expenses and enable faster adaptation to environmental variations. This approach could accelerate adoption of AI systems in manufacturing, logistics, and service robotics where continuous improvement is economically valuable.

The research demonstrates consistent improvements across diverse tasks, suggesting the method generalizes well. Future developments may focus on scaling memory management for long-term deployment, integrating multi-agent experience sharing, and combining retrieval-based learning with lightweight online adaptation mechanisms. The project's public code release indicates potential for broader adoption within the research community.

• →Dejavu enables deployed embodied AI agents to learn continuously from past experiences without retraining frozen policies
• →Experience Feedback Networks retrieve contextually relevant memories and use them to improve action prediction accuracy
• →Post-deployment learning framework reduces computational overhead compared to traditional fine-tuning approaches
• →System demonstrates improved adaptability and robustness across diverse embodied AI tasks in experimental evaluations
• →Architecture combines semantic similarity matching with reinforcement learning to encourage alignment with relevant historical behaviors