OPENING THE LOOP: COMPOSABLE WORKFLOWS FOR SELF-DRIVING LABORATORIES VIA MULTI-AGENT REINFORCEMENT LEARNING

Engineering notes

Across these benchmarks, the RL-SDL reaches target device performance at lower average cost than a fixed-workflow Bayesian optimization baseline while learning to suppress uninformative measurements and preserve device testing for samples with high expected value or high calibration utility.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

Laboratories are open-ended action spaces in which researchers compose experimental workflows from a large set of available procedures, adapting the sequence on a per-sample basis to balance information gain against cost. Self-driving laboratories (SDLs) have demonstrated early success coupling machine learning with automated experimental platforms, but typically operate within fixed closed-loop action sequences that do not replicate the composable, goal-directed nature of real research workflows. Here, a multi-agent framework is introduced that extends this open-ended experimental action space to the SDL. The framework pairs an active learning agent with a reinforcement learning agent capable of dynamically routing samples through a rich set of available experimental procedures. The resulting system exhibits several behaviors absent from conventional SDLs, including dynamic composition of per-sample workflows, the ability to scrap unpromising samples mid workflow, continual learning in which later optimization campaigns benefit from earlier workflow discoveries, and recognition that some experiments are only conditionally informative. The method is evaluated in analytically defined virtual environments designed to isolate three regimes: a single informative surrogate, two complementary surrogates, and four partially informative surrogates whose combined cost exceeds a direct device measurement. Across these benchmarks, the RL-SDL reaches target device performance at lower average cost than a fixed-workflow Bayesian optimization baseline while learning to suppress uninformative measurements and preserve device testing for samples with high expected value or high calibration utility. This work identifies open-ended action spaces as the next frontier in SDL generalization and provides a framework for autonomous experimental design in which the experimental workflow itself becomes a learned output of the optimization.

Links and sources

Comments