ReJ aka Renaldas Zioma @ Telluride’2025
Telluride Neuromorphic Cognition Workshop
Tutorial on Simulation Environments
Challenges in Sim 2 Real
IT IS NOT JUST PHYSICS!
Sim 2 Real
• Physics simulation imperfections
• Sensors —//—
• Imperfect data acquisition
• “Reasoning” integration with the simulation engine
• Real world noise
in sensors & actuators
Physics Engines
• MuJoCo - pretty precise, great for robotics!
• ODE - old, meh
• Bullet - continuous time, partial GPU acceleration, but generally very fast
• PhysX / Havok - fast, supports GPU acceleration!
• DART / Drake - slow, very precise
• BRAX - di
ff
erentiable, supports GPU acceleration!
Complexity levels of physics engines
source: wikipedia
What physics engines actually do?
• Body dynamics (motion)
• Collision detection
• Friction, joints, springs…
• Constraint solver
Body Dynamics
• Usually discreet time steps!
Body Dynamics
• Usually discreet time steps!
• Newton-Euler dynamics
Body Dynamics
• Usually discreet time steps!
• Newton-Euler dynamics
Center of the Mass Velocity, Torque
Collision Detection
• Find if objects overlap
Spatial Hierarchical Representation
Collision Detection for N bodies
• Reduces O(N·N) ~> O(N·logN)
Types of Spatial Hierarchies
Contact Resolution
The simplest case: 2 spheres
Contact Resolution
The simplest case: 2 spheres
Contact Resolution
The simplest case: 2 spheres
Contact Resolution
3 spheres?
Contact Resolution
3 spheres?
Contact Resolution
3 spheres?
Can be solved iteratively
Drag, friction, springs, joints…
Collisions + drag, friction, springs, joints…
Images from: Physics Tutorial - Part III: Constrained Rigid Body Simulation by Nilson Souto
Impulse Based Dynamics
• Collision detection
• Friction, joints, springs…
• Constraints
• Impulses
• Solver
Collisions + drag, friction, springs, joints…
• Collision detection
• Friction, joints, springs…
• Constraints
• Impulses
• Solver
Images from: Physics Tutorial - Part III: Constrained Rigid Body Simulation by Nilson Souto
Another Spatial Hierarchical Representation
Collisions + drag, friction, springs, joints…
Geometry Representations
Primitives
Mesh
Hollow Inside!
Geometry Mesh
Hollow Inside!
Geometry Mesh
Constructive Solid Geometry
Constructive Solid Geometry
Conservative Approximation
Signed Distance Fields (SDF)
Neural Representations
DeepSDF, CVPR2019
Fast Moving Objects
A similar problem to “hollow” geometry
Δt
Δx
Δx
Fast Moving Objects
A similar problem to “hollow” geometry
Δx
Δx
Fast Moving Objects
A similar problem to “hollow” geometry
Δx
Δx
The Future is Differentiable
The Future is Differentiable
• Dynamics
• Geometric primitives
• Geometric meshes non di
ff
erentiable
• Collisions non di
ff
erentiable
• SDF
• Joints
For Physical Simulation
Sensor simulation
Sensor simulation
“Ground Truth” Images
LiDAR simulation
LiDAR simulation
LiDAR simulation
LiDAR simulation
Real Use-Case
2 Hard Problems
Manual Annotation
Near Accident Situations
Scenario Building
Scenes: static and
dynamic elements
constituting stages,
parameters and
actors for the
simulation
Episodes: scene
variations - different
placements and
actions for actors,
scene parameters
(weather, light,
textures)
Scenes: static and
dynamic elements
constituting stages,
parameters and
actors for the
simulation
Episodes: scene
variations - different
placements and
actions for actors,
scene parameters
(weather, light,
textures)
Summary
• Sim2Real Challenges
• Physics and sensor simulation imperfections
• Imperfect and real world data acquisition
• “Reasoning” integration with the simulation engine
• Differentiable physics to close RL loop
• Directions to Bridge Sim2Real Gap
• Domain Randomization
• Learnable Approximations for Complex Real World Dynamics
Reinforcement Learning
54
Reinforcement Learning

Challenges in Sim 2 Real. Tutorial on Simulation Environments.