What is sim-to-real?
Sim-to-real is training a robot's skills in a physics simulator and then transferring them to the real robot, closing the 'reality gap' between the two.
Updated:
Definition
Sim-to-real (sim2real) means learning a behavior — walking, grasping, balancing — in a physics simulator, usually with reinforcement learning, then running that policy on a physical robot.
How it works
Engineers model the robot and its world in a simulator, randomize physics and appearance so the policy cannot overfit, train at massive scale on GPUs, then deploy to hardware and fine-tune with real data. Robot hardware is increasingly designed for it: Boston Dynamics built its 2026 Atlas hand with backdrivable direct-drive joints precisely so it could be simulated with high fidelity (read the story).
Why it matters
Collecting real robot data is slow and expensive. Sim-to-real is one of the main ways legged and humanoid robots learned to walk robustly, and it complements real-world data from teleoperation.
Frequently asked questions
Why train robots in simulation?
Simulation is fast, cheap and safe: millions of attempts can run in parallel, and a fall or collision costs nothing.
What is the reality gap?
Differences between the simulator and the real world — friction, motor delays, sensor noise, object properties — that make a simulated policy fail on hardware.
How do companies close the gap?
Domain randomization (varying physics and visuals in sim), accurate actuator models, and hardware designed to be easy to simulate, as with Boston Dynamics' 2026 Atlas hand.
Coverage
Hesai's Kosmo turns real rooms into robot training worlds in minutes
The lidar maker launched Kosmo One (8K spatial camera) plus Kosmo World cloud, partnering with Galbot and Dexmal; a Dairy Queen scan cut modeling from ~300 hours to 15 minutes.
Galbot's PASSAGE teaches humanoids to duck, step over and squeeze through clutter
On a Unitree G1, Galbot's planner–tracker stack finished 50 unseen physical layouts without falls; 45 were contact-free. Destination success in sim was 98.67% — contact-free 70.3%.
AWS opens a Physical AI Toolchain with NVIDIA stack hooks
On Oct 8, 2026 AWS launched an open-source Physical AI Toolchain that stitches synthetic data, SageMaker training, simulation and edge deploy — with NVIDIA Isaac and Cosmos in the mix — after RoboMaker's 2025 shutdown.
Boston Dynamics gives Atlas a four-finger hand built for mass production
Atlas's new hand has 13 degrees of freedom, no pinkie and direct-drive joints designed for sim-to-real training, as Hyundai starts training Atlas in Georgia and Rohit Prasad becomes CEO.
Skild's S1 learns soccer by playing itself for 140 simulated years
After launching S1 — a robotics foundation model that copies a single video demo without fine-tuning — Skild showed Physical Self-Play: the same base policy trained against itself in Isaac Sim, then transferred to a humanoid.
Apptronik's Robot Park: Apollo 2 fleets collecting data for Gemini Robotics
The Austin humanoid firm expanded its ~90,000 sq ft training hub and wheeled/bipedal Apollo 2 platforms with Google DeepMind, Mercedes-Benz and GXO sites feeding the commercial Apollo 3.
Newsletter
Robrief Weekly
One email a week with the robotics news that matters. Free, unsubscribe anytime.