Simulation can teach robots a lot, cheaply, but contact, friction and rare surprises still push the best systems back to real hardware.
Training a robot in the real world is slow, costly and sometimes dangerous. Training it in simulation is fast and nearly free. So why does almost every serious robot learning effort still collect real data? The answer is that the debate is moving. Newer research shows simulation doing more than before, but the physical world still contains things simulators handle badly. Here is what is proven, who is working on it, and what could change.
How the Technology Works
A simulator is a physics program that stands in for the real world, so a robot can practice millions of times without breaking anything. The trouble is the sim-to-real gap, the drop in performance when a simulated policy meets real hardware. One 2026 explainer splits the gap into four causes: visual differences, physics approximation error, sensor noise mismatch and missing rare scenarios.
The physics cause matters most for manipulation. Researchers note that the gap is largest for tasks involving contact, where interactions are complex and discontinuous. Friction, squishy objects and slipping grips are hard to model. One industry write-up argues the gap is mostly a physics problem rather than a rendering problem. Edge cases are the other issue: rare events a robot meets in deployment, such as a dropped item, a glare or an odd object, that a designer never thought to simulate.
Who Is Involved
Physical Intelligence argues that simulated data alone cannot prepare models for the real world, and builds on actual robot experience. The other side is serious too. The Allen Institute for AI released MolmoBot, an open manipulation model suite trained entirely on simulation data. Academic groups are testing hybrids. One approach fine-tunes a simulator-trained model with a small amount of real tactile interaction data. Real-world datasets are large: Open X-Embodiment combines over one million real robot trajectories from 22 robot types, and DROID has 76,000 teleoperated trajectories.
Separating Evidence From Promise
Working products versus demonstrations. I found no commercial robot sold as trained purely in simulation for contact-heavy work. Simulation-only results are papers and open releases. One data vendor claims every commercial robot learning deployment today uses primarily real data, but that company sells real-world data, so treat the claim as an interested party's view.
Research prototypes versus commercial systems. Simulation-first results are prototypes. A humanoid study reports that policies trained on its simulation data beat those trained on real teleoperation data on most tasks, helped by wider lighting variation. A lecture summary of that paper says a mix of simulated and real data performed best. Ai2 says MolmoBot outperforms π0.5, which was trained on real demonstrations, on pick-and-place benchmarks. Those are benchmark tasks, not production work.
Company claims versus verified results. Ai2 and the humanoid authors report their own results. I found no independent replication. A data vendor also reports that 500 real demonstrations improved success by 18 percentage points over a synthetic-only baseline, but this comes from a seller's summary of a paper, so check the paper itself.
Near-term versus long-term. Near-term, the practical recipe is a blend: co-training that mixes real and simulated data, then fine-tuning on real demonstrations from the target site. Long-term claims go further. One source lists the idea that synthetic data will make real-world collection obsolete as a myth, arguing the contact-rich gap has not narrowed much in five years. That is an opinion, and the evidence above shows simulation is improving, especially for locomotion and pick-and-place.
Why It Matters and What Could Change
The data bottleneck decides who can build capable robots. Real teleoperation needs hardware, space and operators. If simulation closes more of the gap, Ai2 says academic labs without large-scale teleoperation setups could join in. If contact physics stays unsolved, data-rich companies keep the lead.
Watch for three things: independent tests of simulation-trained models on contact-heavy tasks, better contact models in simulators, and published failure rates on rare events. None of these is settled today.