Sim-to-Real Transfer for Robotic Control

Abstract

This project explores the challenge of transferring learned robotic control policies from simulated environments to real-world applications. Through a 15-day intensive research collaboration at KAUST, we developed and validated novel techniques for domain adaptation using reinforcement learning algorithms. Our approach demonstrates successful deployment across six different trajectory-following tasks with a 98% success rate, bridging the sim-to-real gap through systematic domain randomization and robust policy training.

15 Research Days

6 Task Scenarios

98% Success Rate

10k+ Training Episodes

Methods

Training Environment

Physics Engine: IsaacSim 4.5
Algorithm: PPO with domain randomization
Observation Space: 24-dimensional state vector
Action Space: Continuous control signals
Training Episodes: 10,000+

Real-World Deployment

Computing:On-board Computing
Framework: ROS2 Humble
Control Frequency: 100 Hz
Sensors: MoCap (Motion Capture System)
Communication: WIFI

Robot Platforms

Simulation Platform

Environment	IsaacSim 4.5
Control Rate	100 Hz
Sensors	Absolute Position, etc.

Physical Platform

Hardware	CrazyFile
Control Rate	100 Hz
Sensors	MoCap (Motion Capture System)

Experimental Results

Performance comparison between simulation training and real-world deployment across all tasks.

Task Performance Summary

Task	Success Rate	Convergence Time	Tracking Error
AR Action Stabilize	98%	2.3s	-
AR Action Stabilize 2	96%	2.5s	-
AR Observation Stabilize	97%	2.1s	-
Circle Following	-	-	0.03m
Spiral Following	-	-	0.04m
Figure-8 Following	-	-	0.05m

Demonstration Videos

Side-by-side comparison of simulation training and real-world deployment.

Task

Status

Simulation

AR Action Stabilize

Stabilization task using action-based augmented reality control. Success Rate: 98%, Convergence Time: 2.3s