Try Pro free

The best 10 min of 22

Robots Will Train Mostly in Simulation by 2027 | Neil Nie

Neil Nie predicts robots will shift toward simulation training within a year and says local models will handle fast, safety-critical control.

MTS · Sophia Dew

The best minutes

  1. 3:31Robotics can inherit useful knowledge from non-robotics dataPro pick 4 min
  2. 8:14Robots need local models for fast, reliable controlPro pick 6 min

Every moment plays free. PRO marks a write-up that's in Pro.

Also worth watching

  1. 15:11V0.5 combines full-body control with task-state awareness2 min

Skip the 3-minute company and open-weight discussion, the model-leadership exchange, the product update, and the closing sponsor reads.

Full analysis · 6 min read

Real Alpha · Nov 15

Neil Nie expects more robots to train in simulation within one year

  • Nie expects a lot of robots to do more training in simulation within one year.
  • That is a forecast without a baseline, adoption share, or named task.
  • Teams should measure simulated training against physical validation and deployed results.

Free with an account

Read the rest of the analysis

7 sections, the 3 other voices it tests Nie against, and what it means for you. Every episode we cut gets one.

Signed up before? The same buttons sign you in. The episode, its moments and the transcript stay free.

17:54

Nie forecasts more simulated training, without a share

Nie forecasts a directional change, not an adoption rate. He expects a lot of robots to do more training in simulation within one year.

He does not give a share, starting baseline, named task, or measure of training volume. A count of robots and a count of training hours would answer different questions.

One measure would not imply the other. A company could use simulation extensively while still collecting much of its data from physical robots.

Nie also describes a real-to-sim-to-real loop, but does not explain its mechanics. I read that return to real as a validation question, not a claim that physical testing becomes unnecessary.

Our read

The forecast is specific in direction but not scale. A team needs to choose a unit before it can test whether the prediction holds.

18:54

His commercial exposure calls for evidence, not dismissal

Nie is Verne Robotics' CEO. The company released an open-weight world action model for industrial robots.

He said Verne is already deploying with a number of customers and expanding its team across engineering and deployment. He also described deployment quality as the company's north star.

Nie said Verne wants its model to remain commercially available for customers to train and deploy. That puts his forecast beside a product and market he wants the field to adopt.

This context matters, but it does not invalidate his operating experience. The episode does not report a simulation-training share from Verne's customer deployments.

His view comes from a company building and deploying robots, not a market-wide survey. The distinction limits what the forecast can establish.

Our read

Treat this as an informed founder's forecast, not measured industry adoption. Verne's deployments could become useful evidence if their training mix is reported.

18:00

Simulation needs physical checks before it informs deployment

Nie's real-to-sim-to-real idea needs a physical comparison before it becomes operational evidence. Training volume alone cannot show whether a behavior transfers to the actual robot.

Josh Gruenstein, Tutor Intelligence's CEO, said on MTS in November 2026 that Astra can build robot-like simulation environments quickly. He also said real robots remain essential for validation.

That supports the tooling case, not Nie's forecast about how much training will move into simulation. Faster environment creation does not establish deployment success.

For example, a team can compare task results in simulation with the same work on its physical robot. It can then track whether the results persist across model updates.

This test connects simulated training to the thing operators need, reliable performance on actual equipment. The episode supplies no such comparison or result.

Our read

Simulation may expand the number of conditions teams can test. The return to physical robots remains the check on whether those tests transfer.

17:58

Other data problems frame, but do not test, the forecast

Two other robotics conversations add context about data. Neither tests whether Nie's one-year simulation forecast will come true.

Jason Ma, Dyna Robotics' co-founder and chief scientist, said on Latent Space in October 2026 that robotics data quality is non-stationary. His point was that what counts as high-quality data changes as a model improves.

That does not mean older data necessarily loses value. It means teams may need to reassess data quality as their models change.

Ali Ansari, micro1's CEO, said on Sabrina Halper Show in October 2026 that the physical world has a larger action space than digital work. He also said robotics needs broad human-collected data.

These are neighboring constraints, not counterevidence to simulation. They suggest pilots should specify which actions they cover and which data sources supply them.

Take Nie's shelf-stocking example. The robot bends, picks up inventory, places it on a shelf, and turns around.

That is a concrete sequence, but the episode gives no figure for how much of it simulation can cover.

Our read

Ma and Ansari describe data challenges that sit alongside Nie's forecast. They do not show that simulation will, or will not, take a larger role.

10:17

Local policies add a separate deployment constraint

Simulation changes training inputs, while local inference determines whether a robot can act on time. Nie said a robot's model has to produce actions five times per second.

He compared that rate with a human reaction time of about 200 milliseconds. The word about matters because this is his explanatory comparison, not a benchmark result.

Nie said local policies can address latency and Wi-Fi problems. He described larger models running in the cloud every second or even more slowly.

The host's weak-Wi-Fi scenario illustrates the deployment constraint, and Nie agreed with the local-cloud split. For example, a robot in a poor connection area still needs responsive control.

A simulated result that depends on cloud round trips for fast actions would not answer the local-response question. That is an operational test, separate from counting training data.

Our read

The training mix is only one part of readiness. Teams also need to test response time and task completion under real network conditions.

12:48

General models may absorb specialized robotics advantages

Nie's account leaves room for several routes to robot capability. He said GPT-6 Astra surprised researchers by controlling a robot without much robotics-specific training.

He linked that transfer to learning about 3D space, including 3D control and video games. That is one reason broader training data may matter.

On benchmarks, he said a general model can do better in unseen environments. Specialized robotics models can still win on cycle time, speed, and similar measures.

But Nie did not expect that advantage to define the long-term picture. He said ChatGPT could absorb those tasks too.

So the simulation question is one input among several paths to capability. Simulated training volume alone would not tell a buyer whether a robot handles unfamiliar tasks or meets cycle-time needs.

Our read

Nie sees current strengths on both sides, while expecting general models to improve. Buyers should compare generalization and speed rather than treating training location as a proxy for either.

17:10

Verne's model features do not prove the forecast

Verne's model description is about control and task sequencing, not simulation-derived performance. Nie said it supports full-body control for robots moving through factories or warehouses.

He also said the model tracks whether a subtask is complete, allowing GPT-6 Astra to move to the next step. That places the model inside an agentic workflow.

This matters because many deployments involve a sequence of actions, not one isolated motion. But the transcript reports no task-success rate, cycle-time comparison, or customer result for the released model.

Nor does Nie report gains tied to simulation training. Open weights let customers inspect and test the model, but that access is not itself evidence of transfer.

Ask for results separated by task, training source, physical validation, and operating speed. Otherwise, model features and the industry forecast remain separate propositions.

Our read

The described capabilities make Verne's model a candidate for targeted evaluation. They do not substantiate the one-year forecast about industry training.

What it means for you

  1. Set a training baseline. Track simulation and real-world data separately by task, using a unit such as hours or examples. Without a baseline, a change in training mix cannot be verified.
  2. Log sim-to-real misses. Record where simulated training fails on physical robots, including task completion and timing. Compare results across model updates.
  3. Test local response. Measure whether deployed policies meet response needs with weak or unavailable Wi-Fi. Tie the test to the robot's actual tasks.
  4. Request task-level results. For Verne's model, ask for full-body task results, cycle time, and physical validation separately. That would distinguish stated capabilities from deployment evidence.

What to watch

  • Whether Nie's one-year forecast produces a measurable shift in simulated training, and how the shift is counted.
  • Whether Verne reports task-level deployment results tied to training source and physical validation.
  • Verne said it expected to launch more products in the coming month.

Send this to the robotics lead setting simulation and deployment budgets.

17:54

Watch the minute where Nie gives his one-year simulation forecast

Written by Real Alpha from the episode's transcript and the other moments it links, then checked against the transcript. Quotes are the speaker's exact words.

Minute by minute

  1. 0:50Open weights can accelerate deployment

    Nie says Verne’s capture device records customer tasks such as grasping packages and assembly for model training. He describes V0.5 as one of the leading U.S.-made open-source foundation models and argues open weights can lower adoption barriers, support U.S. leadership and serve national-security interests.

  2. 3:31Astra changed robotics expectations

    Nie agrees that GPT-6 Astra was a GPT-3 moment for robotics. He says its ability to control a robot without much robotics-specific training showed that knowledge from 3D control and video games can transfer to robotics. He expects fast reactive models and more capable models to work together.

  3. 6:11Deployment relationships create the moat

    Nie says open models let customers inspect and benchmark systems, while much of a robotics company’s competitive moat comes from customer relationships and deployment. He sees integrators using open models to help labor-constrained, mid-sized factories and warehouses increase output as well as reduce costs.

  4. 8:14Safety favors local inference

    Nie says fast robot-control models should run on-device to avoid latency and Wi-Fi disruptions, while larger models can run in the cloud at a slower pace. He describes humanoids processing camera images on a GPU with a model of almost 10 billion parameters and producing actions five times per second; he says about half the companies he saw at a conference ran policies locally.

  5. 12:00Robots will combine local and cloud models

    Nie names NVIDIA Groot and Physical Intelligence’s model as strong local options. He says GPT-6 Astra does better in unseen environments, while specialized robotics models can win on cycle time and speed; he expects general models to absorb more tasks and says NVIDIA remains the category leader in local open-weight models.

  6. 15:11Video model teams are entering robotics

    Nie says solving robotics problems draws on multiple areas of machine learning, but remains robotics-driven. He points to Runway and Black Forest Labs entering the field and says competition can improve open-weight models. He describes V0.5 as a full-body control model designed for agentic systems that can track task completion.

  7. 17:21Simulation could overtake real data

    Nie predicts that within a year many robots will be trained more in simulation than through real-world data collection. He expects a strong real-to-sim-to-real flywheel and says simulation is an area Verne is betting on.

  8. 18:01Verne expects deployment growth

    Nie says Verne plans to launch more products and expand its team across engineering and deployment. He calls 2026 a year for models and research and expects 2027 to be a year of deployment. The company is already deploying with customers and aims to triple that in the coming quarters.

Published Nov 15, 2026.

Transcript · 16 min read

MTS, hosted by Sophia Dew, with Neil Nie.

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

  1. simulation will become more and more relevant in the age of Ingentic robotics. So I think that's an area that we're excited to make a bet on, right? So in one year, a lot of robots will be trained more in sim rather than collecting real-world data. So there's going to be a really strong reel-to-sim-to-reel flywheel. Hello, everyone, and welcome back to MTS. Today, we are joined by Neil Nhi, the CEO of Verne Robotics that just released V0.5, an open-weight world action model for industrial robots. So, Neil, welcome to MTS. Thank you so much, Sophia. I've been a huge fan of your show, and it's great to be here today.

  2. What an honor. Well, it's great to have you as well. I will say, I love guests who bring props. And here I see maybe it's like a robot that's not connected to any intelligence right now. What's going on here? This is a data capture device that we use to help our customers capture data that they can use to train robots. So you can use it to, let's say, grasp some packages, do some assembly, right? And that data goes into our robot foundation model and can bring them autonomy. Robot Foundation model that is all open way and leading on some benchmarks. What can you share there? That's right. We recently released V0.5, which is our state-of-the-art robot foundation model.

  3. It is currently one of the best U.S.-made open-source foundation models out there, and we're quite proud of that. We think that there is a huge future where we have to use open-source models to then accelerate the deployment of robotics. Yeah. I want to get your perspective because today is a very interesting day to talk about the future of open source, especially when it comes to the U.S. leading in open source. So what are your takes here? That's right. Well, we believe that a number of companies such as NVIDIA and Jensen specifically have really pioneered this idea of a U.S.-led open source effort.

Read the whole transcript
  1. everything from large language models that we are seeing to robotics, right? We think that it is not just in the interest of the U.S., but rather a national security question as well. And we think that having open source models not only lowers the barrier of entry for adoption for other companies, right, but also it provides, I think, a really strong technical moat where the U.S. can lead in this specific technology category. So, of course, NVIDIA was leading when it came to the U.S. being a leader in open-weight models, what do you think today's news might mean for the industry? Well, I think we would hope that open-weight models, whether it's from robotics to large language models, will still play a very important role in the ecosystem, right?

  2. And we hope that stewards of these open-weight models will continue to push the frontier, right? But we do think that this puts us in a more precarious place. But I think it's hard to say what's going to happen next, But I'm confident that companies like NVIDIA and Hugging Face will continue to push the frontier. And we are very excited to continue to build on the platforms that they've created. Why precarious? I think there are a lot of competing interests here, as you know, right, Sophia? But once again, I think part of our job as a robotics company that is vertically integrated from models all the way down to deployment is that we're judged by the quality of the automation we can provide for our customers.

  3. right so we want to you know be laser focused on that objective and that is fundamentally the

  4. Neil Nie · 4 min

    “the thing that surprised me and the community the most was the fact that it was able to learn”

  5. Lovable also handles the backend and infrastructure, including managed Postgres, auth, storage, hosting, payments, and more, so you don't have to wire it all up yourself. And through Lovable's MCP server, you can create and deploy your project directly from the agents you already use. Turn ideas into software people love with Lovable.

  6. Neil Nie · 6 min

    “So fun fact, if you play baseball, right, you know the human reaction time is about 200 milliseconds, right?”

  7. Yeah. I know you say the U.S., especially NVIDIA, was leading the frontier and the charge for having more open-weight models. We're hearing this year is the year that the capabilities are good enough for robotics that we're seeing our chat GPT moment, but in the physical world.

  8. given that maybe NVIDIA who knows i'm speculating here i have no idea i'm not sharing any opinions but say NVIDIA isn't able to lead in this domain anymore who do you think will be the leader when it comes to local open weight models that can be purpose-built used used in robotics so i think that's that's a really good question and we want to make sure that Vern is well-positioned to offer a really great open weights platform, right? So we want to be one of the companies that you can lean on to use these kind of open weights local models that you can run inference on the edge, right? And so we definitely hope that, you know, we will be able to commit resources and train and release these models so that we can serve the community and accelerate the deployment of robots. But I think also we are excited to see if other companies that have released models in the past, like Pi, will also join forth and see if they will release models too.

  9. But I think right now, NVIDIA is still definitely the category leader here. Yeah. I mean, we had Runway on yesterday

  10. Neil Nie · 2 min

    “the model actually has an internal understanding of if a subtask or a task has been completed”

  11. So we talked a lot about capabilities that these open weight models will be able to have when it comes to robotics. Wanted to get your prediction. Anything you think will be true in a year from now that isn't true today? Well, that's a great question. So this is partially inspired by the stuff that we discussed with some of my old colleagues at the conference. One is going to be simulation will become more and more relevant in the age of energetic robotics.

  12. So I think that's an area that we're excited to make a bet on, right? So in one year, a lot of robots will be trained more in sim rather than collecting real-world data. So there's going to be a really strong real-to-sim-to-real flywheel, right? So maybe we can check in in a year and see if that's going to be true. Man, we'll have to test it out. Last question I have for you for people who are just following along. What is the latest with Avern? What can people expect to see over the next few months? Great question. So we just released our new open-ways model, And we're excited to launch a few more products in the coming month.

  13. I won't mention exactly what they are, but stay tuned. We're very excited. And we are also growing and expanding the team all the way across engineering to deployment. So if you're interested, please visit our website. And I think 2026 will be the year of models and a lot of this research, right? And 2027 will be the year of deployment. So we're going to see more and more companies deploying robots. We at Vern are already actively deploying with a number of our customers, and we can't wait to triple that in the next coming quarters. Amazing. Well, Neil, I really wanted to thank you for joining us today on MTS and sharing about the latest updates with Vern and what this means for the future of the robotics industry.

  14. Thank you for joining us. Thank you, Sophia. It's such a pleasure to be here. Such a pleasure. A huge thanks to MTS sponsors. security and governance. KongHQ.com. Finally, prove what your AI investment is worth with Lairdyn. Head to Lairdyn.com today to start measuring your AI's bottom line. DB, trusted to run critical applications by 75% of the Fortune 100, the intelligent data platform for the era. Apply today for access to up to $15,000 in Atlas credits, dedicated technical support, and exclusive events and go-to-market opportunities. MongoDB.com. Chargebee, billing and monetization infrastructure for the AI era. chargebee.com.

  15. ClickHouse is the real-time database built to keep up with every query your apps and agents send its way and it now runs postgres observability and ai tooling on one platform learn more at clickhouse.com you ship fast so do attackers don't assume you're secure prove it horizon 3 go hack yourself with ai visit horizon3.ai slash node zero blitzy autonomous software development for enterprise code bases ship 5x faster blitzy.com adquick make your brand a billboard out of home advertising as easy to scale as digital adquick.com arena measuring ai performance in the real world arena.ai support for the show comes from vcx the public ticker for private tech the u.s stock market started history's greatest wave of wealth creation from factory workers in detroit to farmers in omaha anyone could own a piece of the great american companies but today our most innovative companies are staying private longer, which means everyday Americans are missing out. Until now. Introducing VCX, a public ticker for private tech. Visit getvcx.com for more info. That's getvcx.com.

  16. Carefully consider the investment material before investing, including objectives, risk, charges, and expenses. This and other information can be found in the innovations fund prospectus at getvcx.com. This is a paid sponsorship. You

More interviews