Search papers, labs, and topics across Lattice.
LabEvolver is a novel training-free framework designed for wet-lab agents that utilizes episodic memory to enhance safety and groundedness in robotic tasks. By integrating an adaptive perception and planning loop with an outer evolution loop for skill and strategy distillation, LabEvolver significantly improves performance metrics, achieving a 48.2% reduction in pH-regulation completion time and a 60.0% decrease in safety-gate intercepts. Additionally, it demonstrates versatility by increasing the cumulative success rate in ALFWorld from 76.2% to 91.4% across 500 tasks, indicating its potential for broader applications in automated scientific discovery.
Achieving a 60% reduction in safety-gate intercepts, LabEvolver redefines the capabilities of wet-lab agents without the need for traditional training.
We introduce LabEvolver, a training-free framework that equips safe and grounded wet-lab agents with episodic memory from execution experience. LabEvolver couples a state-grounded inner trial loop for adaptive perception, online planning, and safety validation with an outer evolution loop that distills completed trajectories into reusable skill, strategy, and safety experience. On robotic solution-preparation tasks, LabEvolver demonstrates real-world feasibility, reducing pH-regulation completion time and safety-gate intercepts by 48.2% and 60.0%, respectively. On ALFWorld, it further improves cumulative success rate within 20 steps from 76.2% with ReAct to 91.4% over 500 continual tasks, showing generality beyond wet-lab settings. These results support learn-by-doing experience evolution as a feasible path toward closed-loop automated scientific discovery. The project page is available at https://github.com/AndyGao6186/LabEvolver.
