LabEvolver cuts lab task time by 48%, boosts AI agent success to 91.4%

A new arXiv paper introduces LabEvolver, a training-free framework designed to make wet-lab automation safer and more adaptive. Created by Jingya Wang, Yuyang Gao, Liuzhenghao Lv, Yonghong Tian, and Yuyang Liu, the system gives robotic agents episodic memory derived from their own execution experience. Instead of requiring extensive pre-training on domain-specific demonstrations, LabEvolver uses a state-grounded inner trial loop for adaptive perception, online planning, and safety validation, paired with an outer evolution loop that distills completed trajectories into reusable skill, strategy, and safety experience. This lets the agent 'learn by doing' across repeated tasks, improving both efficiency and safety without finetuning.

The results are striking. In real-world robotic solution-preparation experiments, LabEvolver reduced pH-regulation completion time by 48.2% and safety-gate intercepts by 60.0%, demonstrating immediate operational gains. Beyond wet-lab settings, the framework was tested on ALFWorld, a standard embodied agent benchmark for household tasks. There, LabEvolver lifted cumulative success rate within 20 steps from 76.2% with the ReAct baseline to 91.4% across 500 continual tasks. This generality suggests that episodic experience evolution—rather than heavyweight model updates—can be a practical route to closed-loop automated scientific discovery. The paper (arXiv:2607.27690) includes a project page with additional details, but the core takeaway is that training-free memory and self-evolution are becoming serious tools for laboratory robotics and autonomous experimentation.