NED3 Pro: Sim-to-Real Robotic Reaching with SAC+HER: From MuJoCo Training to Real-World Deployment

Together with my student Danial Zafaranchizadeh Moghaddam, we are exploring a sim-to-real robotic reaching pipeline that bridges deep reinforcement learning and physical robot control.

In this work, we implement a goal-conditioned policy using Soft Actor-Critic with Hindsight Experience Replay (SAC+HER), trained entirely in MuJoCo before being transferred to a real robotic platform. The policy learns to generalise across target positions, enabling robust reaching behaviour under varying conditions.

A key aspect of the system is the tight integration between simulation and reality. We deploy the learned controller on the physical robot while maintaining a synchronised MuJoCo digital twin for real-time monitoring and comparison. The framework combines state estimation, goal-conditioned policy inference, and execution within a unified control loop, alongside XY and YZ tracking visualisations for detailed performance analysis.

In simulation, the SAC+HER policy achieves over 90% success rate on the reaching task, establishing a strong foundation for sim-to-real transfer. This setup enables systematic evaluation of tracking accuracy, generalisation, and failure modes, providing insights into the reliability of learned controllers beyond the simulated environment.

#Robotics #DeepRL #ReinforcementLearning #SAC #HER #Sim2Real #MuJoCo #RobotLearning #ControlSystems #MachineLearning

Видео NED3 Pro: Sim-to-Real Robotic Reaching with SAC+HER: From MuJoCo Training to Real-World Deployment канала Learning Plus To Date (A. Zaraki)