This repository explores Reinforcement Learning (RL) for robotics, with a gradual transition from classical reward-based RL to Deep Reinforcement Learning.
The initial focus is on robotic manipulation in simulation, using MuJoCo, where a Google robotic arm (007) learns to pick up a cube purely through reward and penalty signals, without hard-coded trajectories or heuristics.
- RL-based pick-and-place manipulation
- Pure reward–penalty driven learning
- MuJoCo simulation
- Google robot (007) arm
- Limited manipulation task space (for controlled learning)
- Transition from classical RL → Deep RL
- More complex manipulation tasks
- Navigation + manipulation (moving the cube to different locations)
- Integration with perception (vision-based observations)
- Benchmarking different RL algorithms
- Python
- MuJoCo
- Reinforcement Learning frameworks (to be extended)
- Robotics-focused simulation workflows
This project is part of a long-term exploration into:
- Learning-based control for robots
- Scalability of RL in real-world robotics
- Bridging manipulation and navigation under a unified RL framework