How will robots actually understand a wet-lab protocol?
Today, most lab robots are programmed for a single, fixed task. The big shift coming is the ability to take a free-form biological protocol—written for human researchers—and turn it into a sequence of robot-executable actions. This is the core of protocol grounding. For example, ProtoAct uses a structured framework that parses the text, checks for missing or inconsistent steps, and maps the refined procedure into constrained JSON function sequences [3]. This means a robot could, in principle, read a cell-culture protocol and know exactly what to do, including monitoring conditions and subtasks.
The challenge is that protocols are often implicit: they assume the researcher knows when to wait, what to check, and how to adapt. ProtoAct addresses this by retrieving annotated examples and using a checker to catch gaps. In their benchmark, they manually annotated 22 cell-culture protocols into 258 monitoring conditions and 910 executable subtasks, showing the level of detail needed [3]. This is not just about parsing text; it's about making the robot 'state-aware'—knowing what to do when conditions change, which is critical for real lab work.
Will robots get better with practice?
Yes, and the gains are substantial. Two complementary approaches are emerging: simulation-based augmentation and experience-based learning. Pipette, a simulation platform, allows robots to practice tasks in a virtual lab. With only 30 human demonstrations per task, a vision-language-action model (SmolVLA) improved its success rate from 40.4% to 71.8% after simulation augmentation—a 31.4 percentage-point jump [1]. This is a huge leap, meaning robots can learn complex tasks with very little real-world data, which is often scarce and expensive to collect.
LabEvolver takes a different route: it lets robots learn from their own execution history. In real-world solution-preparation tasks, it reduced pH-regulation completion time by 48.2% and safety-gate intercepts by 60.0% [2]. This means the robot got faster and safer over time, essentially learning by doing. The two approaches are complementary: simulation provides a safe, scalable way to generate training data, while experience evolution allows the robot to adapt to the specific conditions of its own lab. Together, they point to a future where robots improve continuously, not just at deployment.
What about safety and adapting to unexpected events?
Safety is a major concern in wet labs, and the papers show that grounding can actually improve it. LabEvolver explicitly includes a safety-validation loop that intercepts unsafe actions, and its 60% reduction in safety-gate intercepts suggests that as the robot learns, it becomes less likely to make mistakes that require intervention [2]. This is not just about avoiding accidents; it's about building trust so that researchers are willing to let robots work autonomously.
Adaptability is another key area. BioMARS, a multi-agent system, uses a hierarchical architecture where a 'Biologist Agent' synthesizes protocols, a 'Technician Agent' translates them into robotic code, and an 'Inspector Agent' monitors for anomalies [4]. This allows the system to handle dynamic lab conditions and errors, which are common in real experiments. In cell passaging and culture tasks, it matched or exceeded manual performance in viability and consistency [4]. This suggests that grounding isn't just about following a script; it's about building a system that can perceive, reason, and react—making it suitable for complex, variable biological work.
About These Sources
This answer is built on 4 studies (all preprints) — published from 2025 to 2026, 4 from 2024 or later — selected as the most relevant from 4 studies that passed quality screening, drawn from 57 papers retrieved from a database of over 500 million.
Sources used in this answer
Pipette: An Embodied Simulation Platform, Benchmark, and Data-Efficient Augmentation Framework for Wet-Lab Robotics
Pipette, a simulation platform with data augmentation, improved SmolVLA's success rate from 40.4% to 71.8% and pi0 from 37.3% to 44.1% using only 30 demonstrations per task, across a 12-task benchmark.
LabEvolver: Training-Free Experience Evolution for Safe and Grounded Wet-Lab Agents
LabEvolver, a training-free experience-evolution framework, reduced pH-regulation completion time by 48.2% and safety-gate intercepts by 60.0% in real-world solution-preparation tasks, and improved cumulative success rate on ALFWorld from 76.2% to 91.4% over 500 tasks.
ProtoAct: Turning Wet-Lab Protocols into Embodied Robotic Actions
ProtoAct converts free-form protocols into state-aware action sequences, using a manually annotated benchmark (BioP2E) of 22 cell-culture protocols with 258 monitoring conditions and 910 executable subtasks, and demonstrates successful execution in simulation and real robots.
BioMARS: A Multi-Agent Robotic System for Autonomous Biological Experiments
BioMARS, a multi-agent robotic system, autonomously conducted cell passaging and culture tasks, matching or exceeding manual performance in viability, consistency, and morphological integrity, and outperformed conventional strategies in differentiation tasks.