Sim-to-real gap, domain randomization, world models, and virtual commissioning.
Synthetic environment scaling and digital twin precision for zero-error commissioning in production environments. As humanoid robots move from controlled demonstrations toward industrial applications, simulation and digital twin technologies become essential for validating behavior, reducing commissioning time, and ensuring safety before physical deployment.
This session addresses one of the most critical challenges in industrial humanoid robotics: the sim-to-real gap — the performance degradation that occurs when control policies or AI models trained in simulation are deployed on physical robots. Documented performance drops range from 20 to 50 percentage points in real-world tasks.
The panel brings together practitioners from Dassault Systèmes, twinzo, and ROBOTOP — three organizations at the forefront of industrial simulation and digital twin deployment for humanoid robotics.
The sim-to-real gap refers to the difference between how a robot behaves in simulation versus how it behaves in the real world. This gap exists because simulators approximate reality with simplified physics models, idealized sensor behavior, and pre-defined failure modes. For industrial humanoid robots — where precision, safety, and reliability are non-negotiable — closing this gap is essential.
Fingertip deformation, thread engagement, friction transitions (static to kinetic), fabric bunching — none of which simulators model accurately.
Impact: Precision assembly, cable routing, screw insertion
Rolling-shutter smear, blown highlights, autofocus hunting, depth sensor dropouts on reflective surfaces.
Impact: Vision-based manipulation, quality inspection, bin picking
Unpredictable human actions, handover variations, hesitation, context-dependent decisions.
Impact: Human-robot collaboration, shared workspaces
Hardware degradation, unexpected environmental changes, sensor drift over time.
Impact: Long-term autonomous operation
Equivalent actuator models replicate real joint dynamics. By performing system identification on physical robots — analyzing step response data and computing ground-truth joint torques — researchers estimate PID parameters that mirror real joint behavior in simulation. This enables zero-shot deployment of reinforcement learning policies without fine-tuning.
AI policies are exposed to thousands of randomized parameters during training — gravity, friction, joint wear, lighting conditions, object positions — forcing robustness to uncertainty. Used by OpenAI to train a robotic hand to solve a Rubik's cube.
Photorealistic scene reconstruction from real environments. VR-Robo uses 3D Gaussian Splatting to create photorealistic and physically interactive digital twin environments from multi-view images, enabling RGB-only sim-to-real policy transfer.
Unlike static simulation, a digital twin maintains a continuous connection between physical and digital worlds. Applications include virtual commissioning, real-time monitoring, predictive maintenance, and zero-error commissioning.
World models learn to predict the physical consequences of actions. Instead of memorizing visual-motor mappings, they develop an internal understanding of physics — enabling generalization to novel objects and situations.
Brings 3D UNIV+RSES, a virtual twin platform integrating knowledge, know-how, and virtual companions for industrial humanoid robots. NVIDIA Physical AI libraries integrated into virtual twins of global production systems.
View Company ProfileProvides a 3D digital twin analytics platform creating a "spatial nervous system" for industrial sites. Real-time monitoring of assets, AGVs, and personnel. Reduces breakdowns by ~20%, implementable in two weeks.
View Company ProfileModular, open, internet-based platform for robot applications. Simulation techniques for validating robot configurations. Standardized processes within ERP systems for integrated factory automation.
View Company ProfileWe are actively seeking additional speakers with demonstrated experience in:
Of the sim-to-real gap and its specific manifestations in industrial humanoid deployment.
To actuator modeling, domain randomization, and real-to-sim reconstruction for humanoid robots.
From Dassault Systèmes, twinzo, and ROBOTOP for virtual commissioning and digital twin deployment.
Moderated by Dr. Christian Becker-Asano · Main Stage
Patrick Heimburger · fruitcore robotics · Innovation Stage
All sessions, tracks, and keynotes
January 20, 2027 · 14:15–15:00 · Main Stage · Midnightbazar Munich