EnvBridge: Bridging Diverse Environments with Cross-Environment Knowledge Transfer for Embodied AI

TL;DR

The proposed EnvBridge approach alleviates environmental constraints, offering a more flexible and generalizable solution for robotic manipulation tasks, and demonstrates that LLM agents can successfully leverage diverse knowledge sources to solve complex tasks.

Abstract

In recent years, Large Language Models (LLMs) have demonstrated high reasoning capabilities, drawing attention for their applications as agents in various decision-making processes. One notably promising application of LLM agents is robotic manipulation. Recent research has shown that LLMs can generate text planning or control code for robots, providing substantial flexibility and interaction capabilities. However, these methods still face challenges in terms of flexibility and applicability across different environments, limiting their ability to adapt autonomously. Current approaches typically fall into two categories: those relying on environment-specific policy training, which restricts their transferability, and those generating code actions based on fixed prompts, which leads to diminished performance when confronted with new environments. These limitations significantly constrain the generalizability of agents in robotic manipulation. To address these limitations, we propose a novel method called EnvBridge. This approach involves the retention and transfer of successful robot control codes from source environments to target environments. EnvBridge enhances the agent's adaptability and performance across diverse settings by leveraging insights from multiple environments. Notably, our approach alleviates environmental constraints, offering a more flexible and generalizable solution for robotic manipulation tasks. We validated the effectiveness of our method using robotic manipulation benchmarks: RLBench, MetaWorld, and CALVIN. Our experiments demonstrate that LLM agents can successfully leverage diverse knowledge sources to solve complex tasks. Consequently, our approach significantly enhances the adaptability and robustness of robotic manipulation agents in planning across diverse environments.

Figures (8)

• Figure 1: Overview of EnvBridge. In the method for generating code for robot operation, the successfully generated code in the source environment is stored in memory and utilized for code generation (Re-Planning) in the target environment. At that time, codes with high query similarity are retrieved from memory and converted to match the style of the target environment, thereby facilitating smooth knowledge transfer.
• Figure 2: The process flow from Knowledge Transfer to Re-Planning. From left to middle: Using a code example from the target environment as a reference, convert the code from the source environment. From middle to right: Use the target environment's example and the transferred code to generate new code for the task.
• Figure 3: Task-specific and average success rate(%) on RLBench.
• Figure 4: Task-specific success rate(%) on MetaWorld
• Figure 5: Performance comparison across different methods on RLBench.