When AI Enters the Lab: Who Turns Scientific Hypotheses Into Real Answers? | Yunqi Podcast Attent!on

云启资本·November 5, 2026·0·16

From TALOS, a physics AI scientist, to its joint exploration of autonomous laboratories with Danaher, HeTan Intelligence is pushing AI from the digital world into real scientific research.

Over the past few years, AI has made remarkable progress in the digital realm. From molecular design and drug discovery to the generation of scientific hypotheses, AI keeps expanding the boundaries of human exploration into the unknown. But in scientific research, coming up with an idea is often just the first step.

Can a molecule actually be synthesized? Do its properties match expectations? Can experimental results be reliably reproduced? These questions ultimately come back to the lab, answered through real operations and validation, again and again.

In a sense, the better AI gets at computation, the more the experimental execution bottleneck stands out.

Take drug discovery. Scientists might spend only a fraction of their time designing reaction routes; what truly consumes enormous effort is the subsequent rounds of experimentation, sample processing, and purification. Even when a model has produced an answer worth validating, there's still a mountain of work in the lab that must be done by human hands.

So what if AI could do more than just think and decide — what if it could walk into the lab and carry out experiments itself?

This is the question HeTan Intelligence has been exploring.

This March, HeTan Intelligence's physical AI scientist, Talos, spent 50 consecutive days running in a real lab at a leading pharmaceutical company, completing over 8,000 physical operations with zero errors. Behind these numbers lies an important step for AI moving from the digital world into real scientific scenarios.

Recently, HeTan Intelligence also signed a strategic agreement and purchase order with global life sciences company Danaher to jointly advance an AI-robot autonomous bioprocess development lab. This is Danaher's first Embodied Artificial Intelligence Beacon project in China.

In this episode of the Yunqi Capital podcast Attent!on, we invited Zhigang Chen, founder of HeTan Intelligence; Zhuo Fang, Chief Science Officer of Danaher China; and Yu Sang, investor in frontier technology at Yunqi Capital, for a conversation.

From Chen's career across internet, healthcare, and AI, to how Talos completes a real molecular purification experiment; from the limits of traditional lab automation, to how Danaher and HeTan Intelligence are jointly exploring next-generation autonomous labs — this conversation attempts to answer a more fundamental question:

When AI can already propose more and more scientific hypotheses, who will turn them into real-world answers?

What you'll hear

From digital intelligence to physical execution: Why can AI already design molecules, yet experimental validation remains a critical bottleneck in the R&D chain?
The boundaries of lab automation: Why has traditional automation struggled to cover complete experimental workflows? What long-standing problems can robots solve?
Deconstructing physical AI scientist Talos: From sample identification and experimental design to purification and detection, how does AI complete a full molecular purification?
From demo to real-world delivery: Behind 50 days, 8,000+ operations, and zero errors — how does a robot handle complex environments and experimental demands?
The real standards of industry partnership: Why did Danaher choose HeTan Intelligence? Does lab intelligence value speed, stability, or reproducibility most?
The next step for AI for Science: When experimental data flows back into models, can an intelligent closed loop form from molecular design to physical validation?

Guests:

Zhigang Chen — Founder, HeTan Intelligence

Zhuo Fang — Chief Science Officer, Danaher China

Yu Sang — Investor and Executive Director, Frontier Technology, Yunqi Capital

01 AI Can Already Design Molecules

Why Do We Still Need Robots?

The progress of AI for Science in recent years has been exciting.

AI models can help scientists design new molecules, predict molecular structures, and even compress design work that used to take days or a week into minutes. But from a computational result to actually obtaining the target substance, there's still a long road in between.

This was also an important reason Chen decided to start his company.

Before this, he worked at Twitter, Alibaba, Tencent, and WuXi AppTec, consistently focused on using technology to solve real industry problems. During his time as Senior Vice President and Chief Digital Officer at WuXi AppTec, he visited numerous labs and factories, witnessing firsthand how scientists worked at the bench: standing for long periods, hands constantly in motion, eyes observing experimental changes, while simultaneously judging what the data meant and what the next step should be.

AI could already help scientists with some design and decision-making, but the final execution still depended on human time, energy, and experience. Design was only one part of the R&D chain; what truly consumed time was the experimentation that followed.

Chen mentioned in the interview that a client once asked him: if AI only solved 10% of a medicinal chemist's working time, could it go further and address the remaining 70-80% of experimental work?

This feedback made him realize that experimental execution was the more worthwhile area to explore deeply.

In 2022, Yunqi Capital decided to invest in HeTan Intelligence. At the time, ChatGPT had not yet been released, large models were not an industry hotspot, and Embodied Artificial Intelligence was far from receiving today's level of attention. But the team was already exploring how to use AI and robotics to push beyond the boundaries of human execution capability.

By 2023, as VLM, VLA, and other technologies gradually developed, Chen became even more confident in the robotics direction.

Moving from model design to physical execution, from the digital world to real labs, HeTan Intelligence gradually formed its current product direction.

Discussion excerpts

Yu Sang: The AI industry was still in a relatively low period at the time. Why did you decide to go all-in on AI-driven data production and robotics?

Zhigang Chen: Human time, energy, and working capacity all have boundaries. I've always been thinking about how to use new technology to push execution capability forward and break through human limits.

From ChatGPT's emergence at the end of 2022, to VLM in 2023, to VLA in the second half of 2023, robotics suddenly brought a very clear signal. So by the end of 2023, I decided robotics was the direction we should move toward.

Yu Sang: We saw a large number of startups working on models at the time, but molecular synthesis was often the bottleneck. The designed molecules were either extremely difficult to synthesize or required very long cycles. Wet experiments were actually the key position in the entire chain, and also the infrastructure for generating data and feeding it back to AI.

02 In the Lab

How Much Work Still Can't Be Automated?

Lab automation is not a new concept.

In drug discovery, from small-molecule screening to large-molecule research to drug screening, automated equipment has been in use for many years. But why has traditional automation never covered all experimental workflows?

Zhuo Fang, Chief Science Officer of Danaher China, offered observations from the industry side.

First, consistency of experimental results. When scientists perform experiments themselves, even identical experiments can yield different results due to variations in operator, habits, and specific conditions. Sometimes, only a particular experienced colleague can complete an experiment stably;换人 and it becomes hard to reproduce.

This not only affects efficiency but also complicates interpretation of results: was there deviation in the procedure, or is there a problem with the experiment itself?

Second, the spatial and workflow constraints of traditional automation. Many experiments cannot be completed within a single fixed workstation. Samples may need to be transferred between different equipment, and instrument operations are required. As experimental workflows grow more complex, traditional automation often demands concentrating large amounts of equipment in one space, imposing high requirements on lab area and environmental conditions.

For many pharmaceutical companies and biotechs, this is impractical. More complicated still, different labs have inconsistent equipment, layouts, and operating environments. Some labs are even distributed across different floors, requiring cross-space sample transfer — all beyond the scope of traditional fixed automation equipment.

Embodied Artificial Intelligence offers another possibility.

Robots can move within lab environments, understand their surroundings through visual perception, operate different instruments and vessels, and adjust subsequent actions based on experimental results. They don't require all equipment to be rearranged around a fixed workstation; instead, they have the potential to adapt to existing lab workflows. However, truly entering a lab means more than just being able to grasp, move, and tap screens. Experiments have strict operating protocols, and sample handling, cleanliness, instrument use, and result consistency must all meet professional standards. This is an important distinction between lab robots and general industrial scenarios.

Discussion excerpts

Yu Sang: Traditional automation has been around for many years. Why still can't it cover all experimental workflows?

Zhuo Fang: Many processes and steps can't be solved by full automation alone. They require transferring samples from one workcell to another, plus some instrument operations. We can't deploy large numbers of robotic arms to complete all operations — the cost would be uncontrollable. Additionally, as experiments grow more complex, traditional automation needs to put more instruments and equipment in the same space, but many pharmaceutical companies don't have such spatial conditions.

Yu Sang: So this isn't a special problem for certain types of experiments, but a common issue in labs. That Embodied Artificial Intelligence can complete operations in more generalized scenarios is highly promising.

03 Talos

Letting AI Both Think and Act

If traditional automation solves repetitive operations in fixed workflows, what HeTan Intelligence wants to do is enable AI to understand experimental tasks and autonomously complete next steps based on feedback during the experiment.

This is why Talos is defined as a "physical AI scientist."

It comprises two important components: an AI molecular brain and physical AI execution capability. The former handles experimental decision-making; the latter transforms decisions into real-world operations. Only with both can experiments move from goal-setting to actual results.

Take small-molecule purification, HeTan Intelligence's earliest entry point. After a chemical reaction completes, what results is typically not pure target product, but a mixture containing target molecules, unreacted starting materials, and byproducts. Scientists need to purify to separate the target product, then judge whether it can proceed to the next reaction or study.

Chen noted that in medicinal chemists' work, reaction design might account for only about 10% of time, while purification could take up 50%. This is an extremely time-consuming and physically demanding step.

And a complete purification experiment actually involves a series of continuous decisions and operations.

Step one: Identify the sample and understand the experimental context.

Scientists place samples awaiting purification on designated handoff racks; the robot retrieves the sample while scanning its code, reading sample information and understanding its corresponding reaction process.

Step two: AI designs purification conditions.

Talos's AI decision engine designs a solvent system based on reaction information, and determines the conditions needed for subsequent thin-layer chromatography (TLC) spotting.

The robot then samples and completes spotting. If results don't meet expectations, the AI model readjusts the solvent system and verifies again, until suitable conditions are found.

Step three: Execute automated column chromatography.

With purification conditions confirmed, the robot installs the sample and silica column onto an automated chromatography system, operates the equipment screen to complete the run, and uses cameras to recognize the purification process and results.

Step four: Detect and confirm target product.

The robot samples from corresponding collection tubes, sends samples into LC-MS for analysis, and confirms whether target product is present.

Step five: Complete collection and post-processing.

With target product confirmed, the robot combines samples from corresponding collection tubes, then transfers them to a rotary evaporator for subsequent processing.

Throughout this process, what matters most is not simply executing pre-set actions in sequence, but that each step requires judging what to do next based on the previous step's results. This means the robot needs both professional experimental decision-making capability and accurate physical operation capability.

The real difficulty is making AI's decision-making capability and the robot's execution capability form a complete partnership.

Chen also described the team's R&D methodology: first deconstruct complex experimental workflows, design clear human-robot handoff rules and robot operating spaces; then use different models to complete atomic operations like grasping; verify entire workflows in simulation environments; and finally deploy and calibrate in real scenarios.

Among model capability, success rate, generalizability, and response speed, the team currently prioritizes success rate and stability. This explains why Talos's development cannot rely solely on improving model capabilities, but also requires extensive engineering, simulation, and real-world scenario accumulation.

Discussion excerpts

Yu Sang: From when a scientist sets a goal to finally obtaining experimental results, what is the hardest and most critical part for Talos?

Zhigang Chen: You'll find that every next step must be based on analyzing and judging the previous step's results, dynamically adjusting what to do next. So first, you need decision-making intelligence, professional judgment capability, design capability, and adaptability. At the same time, these capabilities must be accurately executed by the robot, operating different vessels and equipment in the lab to achieve your goals. The entire process is actually a combination of knowledge and action.

Yu Sang: In such complex, highly generalizable scenarios, how do you ensure the robot can work stably?

Zhigang Chen: We design the entire workflow in detail, run it completely through in a simulated environment. Once it works, we move from sim to real, entering real-world verification, then customer site deployment and calibration. This not only improves workflow efficiency but also dramatically reduces problem complexity and difficulty, ultimately forming a solution that can truly run at customer sites.

From Demo to Real Experiments

What Industry Actually Values

For Embodied Artificial Intelligence, being able to complete a demonstration and being able to run stably in real environments over the long term are two completely different things.

Labs have especially strict requirements for robots.

In HeTan Intelligence's practice, Talos has already completed head-to-head experimental comparisons with experienced scientists at customer sites, achieving equivalent results. This gave the team great confidence: the robot could not only complete tasks, but produce trustworthy experimental results. But Chen also candidly noted that the robot currently lags behind human scientists in some operation speeds, particularly in recovering from failures and handling exceptional situations — areas still requiring continuous accumulation, and where industry collaboration holds significant importance.

Recently, HeTan Intelligence reached strategic partnership and purchase agreement with Danaher to jointly advance an AI-robot autonomous bioprocess development lab, becoming Danaher's first Embodied Artificial Intelligence BEACON project in China.

Their collaboration is not simply putting robots into labs, but starting from real experimental needs to jointly define application scenarios, equipment adaptation, and workflows.

Fang explained that Danaher has extensive life sciences instrumentation and industry experience, and deeper understanding of pharmaceutical customer needs. AI, robot control, and automated workflow design require close collaboration with external partners.

The two sides' connection originally stemmed from customer collaboration with BeiGene. In December 2025, Chen and Fang had in-depth discussions at an industry exchange event at BeiGene's Guangzhou lab, gradually facilitating subsequent cooperation. From initial exchange to project launch, in less than a year, both sides have begun jointly advancing validation of specific scenarios.

What industry collaboration truly tests is whether technology can enter existing work systems and meet labs' real standards. For example, when operating in biosafety cabinets, robots must not only complete experimental actions but meet cleanliness requirements. Human scientists typically wear gloves and spray alcohol for disinfection. After testing, the team found the robot's front gripper also needed disinfection, so they designed special protective devices and sleeves for the robot, enabling it to perform alcohol disinfection according to experimental requirements. These seemingly minor engineering details actually determine whether a robot can truly enter a lab.

As for how to evaluate an intelligent lab system, Fang offered clear criteria:

In early-stage drug discovery, consistency, stability, and reproducibility of experimental results matter more than pursuing speed alone; at the production stage, speed, throughput, and quality need comprehensive consideration. But regardless of stage, reducing scientists' repetitive labor is a shared goal.

For HeTan Intelligence, 50 days, 8,000+ operations, zero errors is an important milestone validation. Next, they need to replicate this capability across more experimental tasks and different customer labs.

Discussion excerpts

Yu Sang: Why did Danaher choose HeTan Intelligence as a partner?

Zhuo Fang: We mainly learned about this situation from BeiGene, understanding that HeTan Intelligence's solution was something customers truly needed. After multiple rounds of discussion with Chen's team, we felt HeTan Intelligence had considerable innovation in Embodied Artificial Intelligence labs, nationally and even globally. This was a reason we chose to collaborate.

Yu Sang: What metrics does lab intelligence value most?

Zhuo Fang: If we're talking about the process development stage in early drug discovery, speed may not be the highest priority, but consistency, stability, and reproducibility of results are more critical. There's a shared aspiration: human scientists don't want to do repetitive work over and over. Everyone hopes to put their time into more innovation.

From Physical Execution to Data Closed Loop

The Next Stop for AI for Science

If we push the horizon one step further, what HeTan Intelligence wants to do is not merely have robots replace scientists in completing experiments. More importantly, they want the experimental process in the physical world to become a data source for continuous AI model evolution.

In the past, AI models could learn from existing data to help scientists design new molecules. But when models enter real labs, a new problem emerges: how to obtain high-quality data that the models haven't yet learned?

Chen has a direct judgment: the data that large models haven't seen yet may be the most valuable data.

Where does this data come from? The answer lies in real labs. Robots complete experiments, obtain real experimental results, then feed this data back to models, enabling them to further optimize molecular design and experimental decisions. This forms a cycle from digital computation to physical execution and back to digital models.

In the interview, Chen also mentioned the team is exploring cooperation models with large model companies: based on model companies'定向 data needs, using robots and AI models in labs to generate real experimental data, even combining HeTan Intelligence's own models with customer models to complete data production for specific tasks.

This means labs, beyond being execution venues for scientific research, could also become important data infrastructure for next-generation AI models.

Of course, this direction still needs gradual validation through specific collaborations and scenarios. HeTan Intelligence's more immediate priority remains focusing on wet experiments and truly doing well with existing technical routes and application scenarios.

For industry players, AI for Science's development is also showing characteristics of different stages. Fang observed that in molecular design, whether small or large molecules, AI has achieved quite encouraging progress and is gradually becoming a routine tool in daily R&D for some pharmaceutical companies and biotechs.

But lab robot scientists remain in a relatively early exploratory stage. Industry is actively validating whether they can run stably and how much help they can provide scientists.

This also means future autonomous labs are not simply replacing existing equipment with robots, but enabling tighter collaboration among AI models, robots, instruments, and experimental workflows.

Finally, both guests addressed a shared topic: if repetitive operations in future labs can all be completed by robots, where will scientists spend their time?

Fang believes robots are not meant to replace scientists, but to free them from tedious, repetitive operations so they can devote more energy to new research directions. In life sciences and drug discovery, vast unknowns still await exploration. Better drug modalities, better targets, better mechanisms of action — all could become important breakthroughs in the next stage.

Chen deeply agreed. He believes AI is giving everyone stronger capability to explore, but what truly matters is whether these capabilities can be brought into the physical world to validate hypotheses that remain unproven. This is the role HeTan Intelligence hopes to play: providing scientists with better tools to let them explore the unknown faster.

Discussion excerpts

Yu Sang: If the experimental execution环节 can one day be taken on by intelligent labs, how will scientists' work change?

Zhuo Fang: We're not trying to use robots to replace human scientists. The predicament human scientists face now is having to do much repetitive labor and tedious operational work. For improving scientific research quality, this work offers limited help, yet it must be done. If labs can run stably, letting robots handle these tedious, repetitive operations, scientists can have more creativity to explore more directions.

Zhigang Chen: Today AI gives every scientist, every person, stronger capability to explore more unknowns. We need to realize this value in the physical world, to validate whether expectations and hypotheses about the unknown are right or wrong. I think what we do is such a tool, enabling people to better explore the unknown world.

Making Good Science Happen Faster

Returning to where this conversation began: AI can already propose more and more ideas worth validating in the digital world, but the true value of scientific research ultimately still must be answered through experiments in the physical world. From model design to real experiments, from robot execution to experimental data flowing back, HeTan Intelligence is exploring a path connecting AI, robotics, and scientific research.

This path is not easy. Robots need to understand complex experimental workflows, adapt to different equipment and environments, and guarantee result stability under strict experimental protocols. Beyond technical breakthroughs, engineering, scenario adaptation, and industry collaboration all require long-term accumulation.

But this is precisely where the significance lies. When scientists no longer need to spend large amounts of time on repetitive operations, when experiments can run with higher consistency and efficiency, research originally constrained by time, energy, and experimental conditions gains opportunity to explore more directions.

Yunqi Capital has consistently taken AI as its main thread, focusing on the complete chain from foundational capabilities to industrial applications. AI for Science is a direction we value highly. Our investment in HeTan Intelligence reflects our appreciation for the team's understanding of lab challenges and their ability to truly connect AI, robotics, and R&D workflows. Going forward, we will continue watching whether these capabilities can be replicated across more tasks and more labs, delivering measurable value to scientists.

As HeTan Intelligence's mission states: Making good science happen faster.

🎧Subscribe to the Xiaoyuzhou "Attent!on" channel for the full episode.