The Gap Between Ideal and Real: A Robot That Aced the Simulation Got Lost in a Real Warehouse
The Case: A digital twin robot with a perfect demo that wasn't real
At one Demo Day review meeting, it was the turn of a startup spun out of a university accelerator. The founder, Zhe (a pseudonym), holds a PhD in robotics. His team builds autonomous warehouse robots: robots that plan their own routes through a warehouse, avoid obstacles, and carry boxes from the shelves to the shipping area.
Zhe proudly explained: "We built a digital twin system for the warehouse. Before a robot sets out on a task, it can run the route in the virtual warehouse first and only go once it's confirmed to work."
He then played a three-minute Demo video. On screen, ten robots moved through the warehouse on clean, efficient routes, running eight hours straight with zero collisions. Several judges nodded.
He added: "We've already received an LOI from a logistics company."
When it was my turn, I asked just one question: "Was this video shot in a real warehouse?"
Zhe paused. "For now... it all runs in a 3D simulation environment."
📘 Startup Glossary | Demo and PoC A Demo shows people what the product looks like. A PoC (Proof of Concept) proves the technical principle can work. Both matter, but both still live in the lab.
📘 Startup Glossary | LOI An LOI (Letter of Intent) means the other party is interested, but has no obligation to pay. Judges give you credit for an LOI, but they won't treat it as an order.
📘 Startup Glossary | Digital Twin A "twin" of a real physical entity built in a computer, such as a factory, a production line or a warehouse. The key is the word "twin": the virtual model must continuously receive real-world data and stay in sync with the physical entity before it can be used for prediction, testing and optimization. If a model was built only from standard specifications and has never received real data, it is closer to a polished 3D simulation than a true digital twin.
What the judges worry about: the Sim-to-Real Gap
That day I wrote in my review: "Practical feasibility has not been validated: no field validation has been carried out, and the Sim-to-Real Gap is not supported by real data, which may raise concerns about practicality. The project is still at a very early stage; we suggest resubmitting once there is an MVP with real-world evidence."
The problem lies in the word "twin." Zhe's team modeled their virtual warehouse on standard warehouse specifications, without data from any real warehouse. In other words, it was a twin without a sibling.
And a warehouse is not a simulator:
In the simulation the floor is clean. In a real warehouse the floor reflects light and has oil stains and dropped packing tape.
In the simulation the boxes are neat cubes. Real boxes get dented, stacked crooked, or left in the aisle.
In the simulation there are no people. A real warehouse has staff rushing to ship orders, forklifts, and Wi-Fi dead zones.
These variables, invisible in the simulation, are why robots "get lost" on site.
📘 Startup Glossary | Sim-to-Real Gap The gap between simulation and reality: a system that performs perfectly in simulation loses performance once it moves to the real world. It is especially common in robotics, autonomous driving and computer vision.
📘 Startup Glossary | TRL TRL (Technology Readiness Level) uses levels 1 to 9 to show how far a technology is from real-world use. Validation only in simulation or the lab usually lands in the middle; only field validation moves you to the higher levels. Many judges carry this yardstick in their heads.
If I were them: the path from PoC to MVP
After the review results came out, my role was to help the team with recommendations. I asked myself: if I were Zhe, what would I do next?
Step 1: Find a design partner.
You don't need to start with a big logistics company. Find a mid-sized warehouse and convince the owner to let you test in one zone during the night shift for two weeks. They get a free automation trial; you get the most valuable thing of all: real data.
📘 Startup Glossary | Design Partner An early customer willing to provide a site, data or feedback and help you refine the product. They are often your first Early Adopter as well.
What if you can't even find a design partner? Step back: borrow data first.
This is the reality for many early teams: warehouse owners worry that robots will damage goods or disrupt shipping, or simply don't want outsiders on site. Don't give up. You can step back and ask them for their current real data, for example:
The actual floor plan and shelving layout
One week of inbound and outbound records and peak order times
On-site photos or CCTV footage to understand how people and forklifts move
Operating logs from existing equipment
Bring this real data into your simulation so the virtual warehouse looks like a real warehouse and gets as busy as a real warehouse. It's still not field validation, but compared with a demo built entirely on assumptions, it's a big step toward the real world. Your digital twin finally has a twin.
More importantly, when you go back with "simulation results based on your warehouse's data" (how much travel distance a route can save, where congestion hits at peak times), what they see is value that matters to them. Many design partnerships start exactly this way.
📘 Startup Glossary | NDA NDA (Non-Disclosure Agreement). Companies often refuse to share data because they fear leaks. Offering to sign an NDA, committing to use only de-identified data, and deleting it after use can greatly lower their guard.
Step 2: Build an MVP and measure three things clearly.
Once you can get on site, the product doesn't need to be perfect, but it must work in the field and measure three things clearly: task success rate, average number of human interventions, and daily maintenance time.
📘 Startup Glossary | MVP An MVP (Minimum Viable Product) is often misread as a "bare-bones version." The point isn't "minimum" but "verifiable": using the fewest resources to answer the single most critical question in a real environment.
Step 3: Face the gap and fix it fast.
Say the success rate in simulation is 98%, and in the first week on site it's only 75%. That's not failure; it's a list of problems. Fix them one by one, test again, and each round gets closer to reality. If "fully autonomous" proves too hard in the field, maybe "human-robot collaboration" will be even more popular.
📘 Startup Glossary | Build-Measure-Learn and Pivot The Lean Startup's "Build-Measure-Learn" loop: build it, measure the results, learn from them, then start the next round. If what you learn shows the original direction won't work, consider a Pivot: adjusting your product direction or target customers.
Step 4: Turn the trial into a paid one.
When your design partner starts relying on your robots, it's time to talk about a Paid Pilot. Going from an LOI to paid validation is the most convincing step in a judge's eyes.
📘 Startup Glossary | Pilot, Paid Pilot and PMF A Pilot is a small-scale rollout in a real setting. A Paid Pilot means the customer is willing to pay for it. When more and more customers come back to buy, renew and refer others, you are getting close to PMF (Product-Market Fit): the market truly needs your product.
Jesse's Take: Being early is nothing to be ashamed of. Skipping validation is what's dangerous
"Very early stage" is not a verdict; it's a location marker. It tells you: you are here, and the next stop is the real world.
If you can't get into a real site right away, don't stand still. Validation isn't all or nothing. There's a staircase in between:
Pure simulation demo → Simulation driven by real data → Small-scale field trial → Paid pilot
Every step up earns a bit more trust from judges and customers. The worst thing isn't moving slowly; it's staying on the first step while writing your proposal as if you've reached the fourth.
What I most want to tell founders like Zhe is this: write the gap between simulation and the field honestly into your proposal."98% success in simulation, 75% in the first week on site. We found three causes and have already fixed two of them." A sentence like that is more convincing than any perfect demo video. It proves you really have left the lab, and that you know where to go next.
With a proposal that has no field data, judges can only trust your promises. Come back with a proven MVP, and they see your capability.
Startup Glossary Summary
Term | In one sentence | What you've proven |
Demo | Showing what the product looks like | It looks feasible |
PoC | Proof of concept | The technical principle works |
LOI | Letter of intent | The other party is interested |
Digital Twin | A virtual model synced with the real world | Simulation is credible only with real data |
TRL | Technology readiness level | How far you are from real-world use |
Sim-to-Real Gap | The gap between simulation and reality | Must be answered with field data |
Design Partner | An early customer who helps refine your product | Someone is willing to give you a site or data |
NDA | Non-disclosure agreement | Gives the other party confidence to share data |
MVP | Minimum verifiable product | It works in a real environment |
Pivot | Changing direction | You learned something from the data |
Paid Pilot | Paid validation | Someone is willing to pay for validation |
PMF | Product-market fit | The market truly needs you |
This case series is adapted from my experience across many reviews and mentoring sessions. Industries, people and details have been rewritten and do not refer to any specific team.
Further Reading
Related posts on JCXO